Phase separation of competing memories along the human hippocampal theta rhythm

  1. Casper Kerrén  Is a corresponding author
  2. Sander van Bree
  3. Benjamin J Griffiths
  4. Maria Wimber  Is a corresponding author
  1. Centre for Human Brain Health, School of Psychology, University of Birmingham, United Kingdom
  2. Research Group Adaptive Memory and Decision Making, Max Planck Institute for Human Development, Germany
  3. Centre for Cognitive Neuroimaging, School of Neuroscience and Psychology, University of Glasgow, United Kingdom

Abstract

Competition between overlapping memories is considered one of the major causes of forgetting, and it is still unknown how the human brain resolves such mnemonic conflict. In the present magnetoencephalography (MEG) study, we empirically tested a computational model that leverages an oscillating inhibition algorithm to minimise overlap between memories. We used a proactive interference task, where a reminder word could be associated with either a single image (non-competitive condition) or two competing images, and participants were asked to always recall the most recently learned word–image association. Time-resolved pattern classifiers were trained to detect the reactivated content of target and competitor memories from MEG sensor patterns, and the timing of these neural reactivations was analysed relative to the phase of the dominant hippocampal 3 Hz theta oscillation. In line with our pre-registered hypotheses, target and competitor reactivations locked to different phases of the hippocampal theta rhythm after several repeated recalls. Participants who behaviourally experienced lower levels of interference also showed larger phase separation between the two overlapping memories. The findings provide evidence that the temporal segregation of memories, orchestrated by slow oscillations, plays a functional role in resolving mnemonic competition by separating and prioritising relevant memories under conditions of high interference.

Editor's evaluation

This pre-registration study used a proactive interference task in combination with MEG recordings on humans to test predictions of a previous computational model postulating that neural representations of competing memories are associated with varied phases of hippocampus theta-band rhythm. Their results confirmed the hypothesis and suggest that reactivations of target and competitor memories indeed occur at different phases of theta oscillations, which is further related to the intrusion effect in behavior.

https://doi.org/10.7554/eLife.80633.sa0

Introduction

Each day is a flow of events that take place at different times but often in overlapping contexts. Unavoidably, many of the stored memories share similar features, and this overlap poses a major challenge for our memory system (McClelland et al., 1995; Norman and O’Reilly, 2003). The present magnetoencephalography (MEG) study investigates the possibility that the human brain uses a temporal phase code to adaptively separate overlapping memories, enabling the targeted retrieval of goal-relevant information.

Competition between similar memories is considered one of the major causes of forgetting (Anderson and Neely, 1996; Underwood, 1957). A prominent case is proactive interference, where access to a target memory is impaired when overlapping information has been stored prior to target learning (Tulving and Watkins, 1974). This impairment is typically ascribed to the conflict arising from the co-activation of competing memories (Kliegl and Bäuml, 2021). On a neurophysiological level, mid-frontal theta oscillations (3–8 Hz) have been identified as a reliable marker of cognitive conflict in general (Cavanagh and Frank, 2014) and mnemonic conflict specifically (Ferreira et al., 2014; Hanslmayr et al., 2010; Johansson et al., 2007; Staudigl et al., 2010). Increased theta amplitudes may facilitate the emergence of phase separation, where populations of neurons transmitting potentially interfering information are segregated in time along a theta cycle (Buzsáki and Draguhn, 2004).

On a theoretical level, computational models assign a central role to the phase of the hippocampal theta oscillation in the ordered representation of multiplexed information (Lakatos et al., 2005; Lisman, 2005; Lisman and Idiart, 1995; Lisman and Jensen, 2013; Nyhus and Curran, 2010; O’Keefe and Recce, 1993). Empirically, in rodents it has been shown that bursts of gamma power reflecting sequences of spatial information processing are represented at different phases of the hippocampal theta oscillation (Bragin et al., 1995; Colgin et al., 2009). Several recent studies in humans suggest that the neural representations of items held in working memory or encoded into or retrieved from episodic memory (Kerrén et al., 2018; Kunz et al., 2019; Pacheco Estefan et al., 2021) cluster at distinct phases of the (hippocampal) theta rhythm. These findings suggest that low-frequency oscillations provide time windows for the selective processing and readout of distinct units of information (Lisman and Jensen, 2013). This study investigates phase separation as a potential mechanism to avoid interference between multiple competing memories that are simultaneously reactivated by a cue.

The computations by which the human brain achieves a separation of multiple overlapping memories are currently unknown. One computational model leverages different phases of a theta oscillation to iteratively differentiate target from competitor memories (Norman et al., 2006). In this model, which we will henceforth refer to as the oscillating interference resolution model, a retrieval cue will activate associated units, representing target and competitor features, in a phase-dependent manner. In the most desirable output state, at medium levels of inhibition, the cue only activates the target units and no or few competitor features. When inhibition is raised towards the peak of the oscillation, only strong features of the target memory will remain active, and the model learns (via contrastive Hebbian learning) to strengthen through long-term potentiation (LTP) the weaker target nodes that did not survive the higher inhibition levels. Conversely, during the transition from a medium to a lower inhibition state towards the trough of the oscillation, activation spreads to more weakly associated units, including some competitor features. This opposite phase is used to identify and punish overly strong features of the competing memory through long-term depression (LTD). The mechanism (see Figure 1) is repeated across several cycles of an oscillation, which changes the similarity structure of memories into a state in which they are less likely to interfere with each other. This model offers a mechanistic but not yet tested explanation for how the brain solves mnemonic competition.

Paradigm and rationale for decoding analyses.

(a) At encoding, subjects (n = 24) were instructed to memorise the word–image associations using an imagery strategy and constantly update their memory with the most recent associate to each word. The experiment consisted of three different conditions. In non-competitive single-exposure condition (NC1), subjects encoded the word together with one associate; in the non-competitive dual-exposure condition (NC2), subjects encoded the word with the same associate twice; and in the competitive condition (CC), subjects encoded the word together with two different associates (one scene and one object). (b) At retrieval, participants were instructed to remember the most recently encoded associate when prompted with a word cue. (c) Subordinate-category classifiers (animate/inanimate for objects, and indoor/outdoor for scenes) were used to obtain independent evidence for target and competitor reactivation at each sample point. Note that the superordinate (object/scene) classifier cannot discriminate between evidence for the target and against the competitor, and vice versa. (d) Hypothesised reactivation dynamics of targets and competitors relative to theta phase, based on the oscillating interference resolution model (adjusted from Norman et al., 2006). Blue circles represent the target memory, pink circles the competitor memory. Target and competitor memories consist of a number of features (small circles), some distinct and some overlapping, that can be either active (dark blue) or inactive (light blue) at any given point along the phase of the oscillation. When inhibition is high, only a few features of the target memory are active initially. The transition from medium to high levels of inhibition is used in the model to identify and strengthen weak target features, such that these nodes have a higher likelihood of becoming active in future high-inhibition phases (see right portion of the graph). In contrast, during the low-inhibition phase, features of both target and competitor memories are simultaneously active. The transition from medium to low levels of inhibition is used to identify and punish overly strong competitor features, which will in turn be less likely to activate even in low-inhibition phases in the future. (e) An illustration of the expected phase of maximum target and competitor reactivation, and their phase difference, in the first and last repetition of retrieval. Note that in this example target and competitor memories show a consistent phase angle across subjects in the third repetition; however, this is not a necessary assumption for finding consistent phase separation.

A previous study (Kerrén et al., 2018) provided proof-of-principle evidence that the reactivation of a single visual memory is rhythmically modulated by the hippocampal theta rhythm. In the present study, we applied the same analytic approach in an AB-AC associative interference paradigm (Figure 1a and b) to investigate the phase modulation of competing memories. Hypotheses for the experiment were pre-registered on OSF. Based on the oscillating interference resolution model (Norman et al., 2006) and related theta-gamma models (e.g., Lisman and Idiart, 1995), we expected that target and competitor memories become active, and can thus be decoded, at different phases of the hippocampal theta oscillation, with only target features active at higher inhibition phases, and competitors surfacing as inhibition ramps down to medium and low levels (see Figure 1d). The model also assumes that across several repetitions of the theta cycle, weak features of target memories are strengthened while overly strong features of competing memories are punished to make them less interfering on future cycles. We hypothesised that the effects of these dynamics will become visible across several repetitions of competitive retrieval and lead to an increasing phase separation because over time target memories become more likely to activate in the high-inhibition phase whereas competing memories require increasingly lower levels of inhibition to activate. Lastly, in line with previous literature (Wimber et al., 2015), the strengthening of target and weakening of competing memories should also be reflected in corresponding changes in their decodability.

Results

Participants (n = 24) completed an associative memory task including one proactive interference condition and two control conditions (Figure 1a and b). In each learning trial, participants were instructed to form a novel association between a unique verb and an image, which could be either an object or a scene. In the critical competitive condition (CC), the word was associated with two images, one object and one scene image, learned on different trials. Participants were asked to continuously update their memory to remember the most recently encoded associate of each word, thus allowing us to investigate the extent to which previously encoded associations interfered with newly encoded information (proactive interference). In the single-exposure non-competitive condition (NC1), a word was associated with only one image, and never reappeared during learning. This condition served as a baseline for behavioural interference effects. In the dual-exposure non-competitive condition (NC2), a word was also associated with one image only, but this association was learned twice. This condition was included as a control for neural (cue) repetition and order effects. Both non-competitive conditions were also used to train and validate our multivariate classifiers. CC trials were pseudo-randomly intermixed with NC1 and NC2 trials during learning and recall blocks.

In a subsequent memory test, participants were asked to remember the most recent image when prompted with the cue word, and cued recall was repeated three times per word–image association, in a spaced fashion, irrespective of condition. Once participants indicated, via button press, that they had the associated picture back in mind, two follow-up questions were asked, one about the supraordinate (object/scene) and one about the subordinate (animate/inanimate for objects, indoor/outdoor for scenes) category. This task design, with its supra- and subordinate categories, is critical because it allowed us to disentangle, on a neural level, the representational evidence for target and competitor reactivation at any given time point, using two independently trained linear decoding algorithms: one to discriminate indoor and outdoor scenes, and one to discriminate animate and inanimate objects (see example in Figure 1c).

Behavioural indices of proactive interference

We first evaluated behavioural evidence for proactive interference, that is, the extent to which cued recall performance suffered from having encoded two different pictures with one word (CC) compared to only one picture (NC1). Only trials where participants correctly responded to both follow-up questions (i.e. they were able to provide the supra- and subordinate category of the target image) were considered a successful recall. A one-factorial repeated-measures ANOVA, with memory accuracy as the dependent variable, and condition (CC, NC1, NC2) as within-subject factor, showed a significant main effect of condition, F(1.453, 46) = 57.99, p<0.05. The main planned comparison of interest showed that memory performance was indeed significantly reduced in the competitive (M = 53.18%, SD = 19.20%) compared to the single-exposure (M = 57.71%, SD = 13.98%) condition, t(23) = 1.80, p<0.05 (Figure 2a). This confirms the first behavioural hypothesis (OSF) that retrieving a memory is more difficult when a previously encoded, now irrelevant associate competes for recall, indicative of proactive interference. For later analyses relating neural and behavioural indices, an interference score for each participant was calculated by subtracting average memory performance in the CC from the NC1 condition. A second post-hoc test showed that a significant difference was also obtained between NC1 and NC2 (M = 78.88%, SD = 10.06%), t(23) = –11.91, p<0.05, demonstrating that, unsurprisingly, a second encoding exposure improved performance compared to a single exposure.

Figure 2 with 2 supplements see all
Behavioural results, time–frequency analysis of theta power, and decoding accuracies in the non-competitive (NC) and competitive conditions (CCs).

(a) As expected, we found that memory accuracy (on the two follow-up questions combined), averaged across the three recall repetitions, was significantly impaired when encoding a given word cue with two different images (CC) compared to just one image (NC1), indicative of proactive interference (red line indicating significant difference at p<0.05, using Paired Samples t-test). Recall performance also benefited from learning a cue word together with the same image twice (NC2) compared to once (NC1). (b) The average intrusion score shows that errors were not random (50% black line), but instead were significantly biased towards the competitor’s subcategory. The proportion of intrusions did not decrease significantly across repetitions. (c) Contrasting oscillatory power elicited by the cue in the CC and NC2 conditions resulted in a significant cluster (500–700 ms, darkest red, cluster-based permutation < .05) in the theta frequency range (3–8 Hz), most prominently over right frontal electrodes (see right inlay). (d) Results of a linear discriminant analysis (LDA)-based classifier trained and tested on the non-competitive conditions (NC1 and NC2) at retrieval, showing a cluster of significantly above-chance decoding accuracy approximately 2500–3000 ms post-cue onset, with an earlier decoding peak around 1–2 s not surviving cluster correction. Graph shows mean and SEM. (e) Results from a classifier trained on the non-competitive conditions (NC1 and NC2) and tested on the CC. Separate classifiers were used to detect target (blue) and competitor (red) evidence at the level of subcategories. No significant cluster emerged when averaging across all repetitions. (f) Realigning the trials to the time of subjective recollection (i.e. response), instead of cue onset, significant target decoding was found in the CC when averaging over all repetitions. (g) Response-locked target decoding in the CC was also significantly higher on correct than incorrect trials. (h) Using a repeated measures ANOVA to con decodability in the first and third retrieval repetition for target and competitor memories, respectively, yielded a significant interaction such that evidence for target memories increased as a function of repetition, whereas evidence for competitor memories decreased as a function of repetition. For behavioural analyses, 24 participants were included, whereas for neurophysiological analyses 21 participants were included. Red line in panel (d, f, g), indicates time windows of significant clusters at pcluster<0.05 corrected for multiple comparisons across time.

Another behavioural index of competition is the intrusion score; a participant’s tendency to choose the competitor instead of the target. Our design allowed us to quantify intrusions based on how often participants selected the specific subcategory of the competitor (i.e. ‘outdoor scene’ in the example in Figure 1) during the staged retrieval. On trials where participants made an error (i.e. selected ‘object’ though the target was a scene and vice versa), they showed a significantly above-chance (i.e. above 50%) probability of selecting the subcategory of the competitor (on average, 88.72%, SD = 10.62%; t(23) = 10.48, p<0.05), suggesting that their errors were not random but biased towards the category of the previously learned but now irrelevant associate (Figure 2b). The proportion of intrusions did not change significantly across repetitions (first: 88.08%, SD = 12.04%; second: 89.23%, SD = 10.62%; third: 88.85%, SD = 14.72%, Z = –54, p=0.59; Wilcoxon signed-rank test of linear slope against zero). Similarly, we did not find a decrease across repetitions in behavioural accuracy for CC target items (Z = –0.73, p=0.47; Wilcoxon signed-rank test of linear slope against zero; see Figure 2—figure supplement 1a). Across participants, the average number of intrusions in the CC correlated negatively with memory performance in the non-competitive conditions, r = –0.37, p=0.037, p<0.05 (see Figure 2—figure supplement 1b), indicating that better memory performance was generally associated with fewer intrusion errors.

The average response time for subjective recollection (indicated by button press) was 2.11 s (SD = 0.15 s) when collapsing across all conditions and repetitions. Response times shortened across repetitions in all three conditions (CC: rep1: 2.29 s, rep2: 2.05 s, rep3: 2.04 s; Z = –2.89, p=0.004; NC1: rep1: 2.55 s, rep2: 2.37 s, rep3: 2.26 s; Z = 3.03, p=0.003; NC2: rep1: 2.01 s, rep2: 1.75 s, rep3: 1.68 s; Z = –4.17, p<0.001; Wilcoxon signed-rank tests of linear slope against zero), suggesting that the time point when participants remembered the associate shifted and overall was faster after repeated recalls, as would be expected.

Mnemonic competition is associated with a frontal theta power increase

Having established that the paradigm elicits competition on a behavioural level, we next checked whether the CC was associated with the neural indices of competition typically reported in the literature. A well-documented effect is an early increase in theta power (3–8 Hz) over frontal sensors, which has been related to cognitive conflict processing in tasks with and without memory demands (Cavanagh and Frank, 2014; Ferreira et al., 2014; Hanslmayr et al., 2010). In this study, the CC and the non-competitive neural baseline (NC2) were equated for repetitions of the cue. When contrasting these two conditions, we expected to see a similar power increase in the theta range in the present experiment. A non-parametric cluster-based permutation test, focused on a time window of interest 0–1 s following cue onset and on the previously reported 3–8 Hz theta range, revealed a significant power increase when retrieving the target association in the CC compared to the NC2 condition (pcluster<0.05), ranging from approximately 500 to 700 ms post-cue (Figure 2c). In addition to behavioural evidence for proactive interference, we thus also replicate previous work that established frontal theta power increases as a neural signature of competition. Note, however, that in terms of its topography, this frontal effect was more right-lateralised than the typically observed mid-frontal theta (Cavanagh and Frank, 2014), and might instead be more related to response conflict or response inhibition (for a review, see Aron et al., 2004). The relationship of this faster frontal to the typically slower hippocampal rhythm (see below) is currently unclear, and it would be an interesting target for future studies to investigate whether the two rhythms relate to a common pacemaker (e.g. the medial septum, see Wang, 2002).

Pattern reactivation of target and competitor memories

The main pre-registered hypothesis of this study concerned the timing of neural reactivation of target and competitor memories relative to the phase of the theta rhythm. As a first step, we thus checked whether and when the categorical content of the associated memories could be decoded from the MEG sensor patterns in the recall phase of the experiment. Two separate linear discriminant analysis (LDA)-based classifiers were trained to discriminate the subordinate picture class, that is, animate vs. inanimate for objects, and indoor vs. outdoor for scenes. As noted above, the two subordinate categories were an important feature of the task design that allowed us to obtain independent indices of target and competitor reactivation. Both classifiers were trained on distinguishing subcategory membership (animate/inanimate and indoor/outdoor for objects and scenes, respectively) in the non-competitive (‘pure’) conditions during recall.

As a sanity check, we then first tested these classifiers’ ability to detect the retrieval of objects and scenes in the non-CCs (NC1 and NC2). Training and testing were conducted time point per time point, in steps of 8 ms, each time point centred in the middle of a Gaussian smoothing window (full-width at half maximum [FWHM] = 40 ms). MEG gradiometer sensor patterns (i.e. signal amplitude on each of the 204 channels) were used as classifier features, independently per participant and per time bin (from –500 to 3000 ms around cue onset). Since the same data was used for training and testing in this case, a tenfold cross-validation was used and repeated five times. Results are shown in Figure 2d, where highlighted in red are significant time points corrected for multiple comparisons over time using a non-parametric cluster-based permutation test (as implemented in FieldTrip toolbox version 2019, Oostenveld et al., 2011). This first decoding analysis indicated that reliable (pcluster<0.05) above-chance decoding performance was found in a time window from 2.55 to 2.92 s post-cue onset, somewhat later than the typical time window for reinstatement of associative memories in cued recall tests (for a review, see Staresina and Wimber, 2019). Earlier clusters did not survive stringent cluster correction. For decoding during visual presentation of the stimuli at encoding, see Figure 2—figure supplement 2a.

To test for reactivation of target and competitor memories in the CC, trials were split into those where the target was an object and the competitor was a scene, and vice versa. The corresponding object (animate vs. inanimate) and scene (indoor vs. outdoor) classifiers, again trained on non-competitive recall trials, were then used to indicate evidence for target and competitor reactivation, restricted to only correct trials. Figure 2e shows the average target decoding accuracy in the CC. Classification performance peaked in time windows similar to the non-competitive condition. However, no significant cluster could be identified when correcting for multiple comparisons across time points, indicating that overall reinstatement was less robust in the condition where two associates competed for recall, potentially because this competition was associated with less confident target retrieval. A more technical explanation for the absence of significant target decoding in the CC is the variance in the timing of subjective recollection across trials and participants, making it difficult for a classifier to detect a consistent time point of neural reinstatement. We thus conducted an additional analysis (not pre-registered, added as part of peer review process) realigning the decoding data to the subjective recollection button press, and thus to the time point when participants indicated they had the associated image back in mind. This method revealed two clusters of significant target classification preceding the button press by approximately 200 ms (Figure 2f). The realigned analysis further revealed that target classification on correct trials was significantly stronger than on incorrect trials (Figure 2g). These results reaffirm that the classifiers are able to pick up categorical reactivation also in the CC, and that this reactivation varies with behavioural performance in a meaningful way. They also suggest that the lack of significant decoding in the cue-locked analyses is due to timing differences in memory reinstatement between conditions, trials, and participants, which is rectified when locking to the time point of subjective recollection.

In addition to general decodability when averaging across all retrieval repetitions, we also predicted in our pre-registration that in the CC the neural representation of the target memories would become stronger and the competing associate would become weaker across repetitions (see Wimber et al., 2015). A contrast of the decoding timelines between the first and third repetition revealed a significant increase (pcluster<0.05, 1.79–2.11 s; third repetition compared to chance pcluster=0.07) in decodability of the target memory approximately 2 s after cue onset (Figure 2—figure supplement 2b). No corresponding cluster emerged when comparing the entire timeline of competitor reactivation between the first and third recall (Figure 2—figure supplement 2c). An additional analysis testing for competitor evidence on incorrect trials, when competitors are more likely to interfere, revealed more robust competitor activation (see Figure 2—figure supplement 2d). Zooming in on the time window where targets and competitors were maximally decodable (1.77–1.93 s after cue onset, puncorr<0.05, see Figure 2—figure supplement 2e), a significant repetition-by-associate interaction was found (F(1,20) = 5.187, p=0.034), with a significant increase of target reactivation from first to third recall (first: 49.43%, SD = 4.08%; third: 53.14%, SD = 6.43%; t(20) = –2.46, p=0.02), and a numerical but non-significant decrease of competitor reactivation (first: 52.05%, SD = 4.55%; third: 50.82%, SD = 5.54%; t(20) = 1.1, p=0.29) (Figure 2h). Thus, we observed a strengthening of the relevant target memory and a numerical but non-significant weakening of the competing memory across repetitions in this select time window. Lastly, taking inspiration from previous work on competitor reactivation and its relationship to behavioural memory interference (Newman and Norman, 2010), we extracted the average decoding across repetitions for the competitor memories from 0 to 1 s after cue onset and correlated this reactivation with the probability of intrusions in the first repetition. This revealed a significant correlation (r = 0.446, p=0.04; Figure 2—figure supplement 1c), such that stronger reactivation of competitor memories positively correlated with a higher level of initial intrusions.

Together, the cue-locked decoding results provide evidence that target memories can be decoded from MEG signals when no competition is present, while under conditions of interference, target reactivation is less robust in early trials but gradually increases over repetitions. This pattern could emerge due to the gradual strengthening of the relevant target memory or due to reduced interference from the competing memory. When instead realigning the decoding to the subjective time of recollection, we found significant target decoding, indicating that temporal variance in memory reactivation could explain the non-significant results in the cue-locked analysis.

Hippocampal theta oscillation clocks the reactivation of target and competitor memories

The next step was to assess whether the fidelity with which target and competitor memories can be decoded fluctuates in a theta rhythm. The analysis was conducted to determine the dominant frequency in the decoding timelines, to replicate our previous findings for target memories (Kerrén et al., 2018), and to extend the principle to competitor memories. We estimated the oscillatory component of the fidelity values, following our pre-registration and the rationale of our previous study (Kerrén et al., 2018). We used Irregular-Resampling Auto-Spectral Analysis (IRASA) (as implemented in the FieldTrip toolbox version 2021), a method that robustly detects and separates the oscillatory from the fractal signal component in electrophysiological data (Wen and Liu, 2016). IRASA was applied to the single-trial fidelity timecourses in the CC, in a time window from 500 to 3000 ms post-cue onset, and separately for the target and competitor memories. The first 500 ms were excluded to attenuate the influence of early, cue-elicited event-related potentials, and because no neural reinstatement is expected during this early cuing period (Staresina and Wimber, 2019). The analysis identified 3 Hz as the frequency with the highest power for both target and competitor memories (same peak frequency was also observed for the first repetition; Figure 3—figure supplement 1a and b), and power at 3 Hz exceeded the 95th percentile of the empirical chance distribution generated from surrogate-label classifiers (Figure 3a and b, red line indicating significant deviation). For the NC, the peak frequency was also in the low theta range but slightly faster (4–5 Hz; see Figure 3—figure supplement 1c and d). Since our theta-locked analyses were focused on the CC, we used 3 Hz as the modulating frequency in all subsequent analyses.

Figure 3 with 2 supplements see all
Rhythmic fluctuations in classifier fidelity and phase locking to hippocampal theta oscillation.

(a, b) An estimation of oscillatory components in the classifier fidelity values, using Irregular-Resampling Auto-Spectral Analysis (IRASA), revealed significant rhythmicity at 3 Hz, compared with the same estimate from a label-shuffled baseline classifier (dashed black line; red line indicates frequencies that significantly differ from baseline) for both target (a) and competitor (b) memories. (c) The hippocampal region of interest used as the source for theta phase, shown in sagittal and (left) coronal (right) plane. (d) Distribution of each individual’s modulation index (MI) with the 95th percentile subtracted. The third retrieval repetition shows significant modulation exceeding the 95th percentile (red line). (e, f) Results of the phase modulation analysis linking decoding fidelity to the phase of the hippocampal 3 Hz oscillation. Fidelity values for target (e) and competitor memories (f) were significantly modulated by the hippocampal theta rhythm at 3 Hz. The histogram shows the chance distribution of the MI, together with the 95th percentile in grey vertical bars and empirical value in blue for target memories and orange for competitor memories. For all analyses, n = 21 participants were included. Graphs in a and b show means and SEM (error bars).

Target and competitor reactivations peak at distinct theta phases

The key hypothesis was that over time the reactivated representations of target and competitor memories would become optimally separated along the phase of the theta oscillation. This hypothesis follows from the oscillating interference resolution model motivating our work (Norman et al., 2006) and other empirical studies on retrieval-induced forgetting (RIF; e.g. Anderson et al., 1994; Wimber et al., 2015). The model assumes that target and competitor representations are overlapping initially, such that strong competitor nodes can (incorrectly) activate during the high-inhibition (‘target’) phase of theta, but gradually get weakened and thus require lower levels of inhibition to become active. Meanwhile, the weakest target nodes do not survive high inhibition initially and thus only activate during a lower inhibition phase of the theta cycle; however, with repeated strengthening they will become active at an increasingly early, higher-inhibition phase. Therefore, while early in time the target and competitor features overlap in their reactivation phase, this overlap is reduced by the strengthening and weakening dynamics in the model. Note that the original model (Norman et al., 2006) predicts a gradual increase of target–competitor separation with each oscillatory cycle of excitation–inhibition. In our pre-registration, we instead hypothesised that the phase separation would be measurable when comparing early with late recall repetitions, allowing us to obtain robust trial-by-trial indices of target and competitor reinstatement and avoid confounds with time-on-trial. As also noted in our pre-registration, there is an influential theta phase model (Hasselmo et al., 2002) suggesting that encoding and retrieval operations in the hippocampal circuit are prioritised at opposite phases of the theta rhythm (for evidence in humans, see Kerrén et al., 2018). Taking this model into account, we thus hypothesised that the target–competitor phase segregation would over time and repetitions become optimal within the retrieval portion (e.g. half) of the theta cycle, rather than spread out across the entire cycle (not shown in Figure 1d).

The phase binning method used to calculate the MI (see above) allowed us to determine the theta phase bin at which target and competitor reinstatement was maximal in a given trial and condition. Importantly, the absolute phase of reinstatement in terms of its angle is likely to vary across participants (e.g. due to individual differences in anatomy). We therefore contrasted the phase of maximal target and competitor reinstatement within each individual participant, computing their phase distance as an index of phase separation. This phase distance is expected to be consistent across participants irrespective of the absolute angle of target and competitor reactivation, and can thus be subjected to group-level statistics. We used a Rayleigh test for non-uniformity (i.e. clustering) to test how coherent the phase separation angle was across participants and a circular v-test to establish whether the mean separation angle significantly deviated from zero. Note that without significant clustering it is difficult to interpret the mean angle in such a phase analysis. On the other hand, significant clustering without a significant difference from zero would indicate that consistently across participants, targets and competitors reactivate at a similar theta phase.

Figure 4a shows the results of this phase separation analysis. We found a consistent phase difference between target and competitor reactivation exclusively in the last repetition (Rayleigh test for non-uniformity, z(20) = 5.03, p=0.005), clustered around a mean angular difference of 34°. Furthermore, a test against zero phase shift confirmed that this target–competitor distance was significantly different from zero difference (z(20) = 8.48, p=0.004). This was not the case for all repetitions averaged, (z(20) = 0.84, p=0.44), nor for the first (z(20) = 2.19, p=0.11) or second (z(20) = 1.76, p=0.17) repetition separately (Figure 4a). Furthermore, there was no significant difference in target–competitor phase separation between the first and third repetition when comparing the difference in mean angle (z(20) = 1.09, p=0.29). Note, however, that there was no significant clustering around a stable mean angle in the first repetition (see above), and this statistical comparison is therefore inconclusive. Finding the separation between target and competitor memories only in the last repetition rather than throughout the entire recall phase deviates from our pre-registered hypothesis, although paralleling the phase modulation results reported above. Overall, the phase distance analysis thus partly confirms our hypotheses, indicating that competing memories become increasingly separated along the hippocampal theta rhythm over time, with significant phase separation emerging after several recall cycles.

Figure 4 with 1 supplement see all
Target–competitor phase difference results.

(a) Phase difference of maximal target minus competitor reactivation for each participant, separately for each recall repetition. At repetition 3 (right circular plot), there was a phase shift of on average 34° that was statistically coherent across subjects and significantly different from zero, indicative of robust target–competitor phase segregation. (b) When splitting the phase difference in the third repetition by behavioural intrusion score, the high-intrusion group had a mean phase separation angle close to zero (top), while the low-intrusion group had a mean separation angle of 57° (bottom). (c) A cross-correlation between the continuous fidelity timecourses of target and competitor memories revealed a significant cluster (pcluster < .05) from approximately 1 s after cue onset, lasting to the end of the trial, with a maximum around 30° phase lag. z-values in panels (a, b) (colour code) indicate strength of Rayleigh test for non-uniformity, with z-values/colours exceeding the dashed line on the colour bar indicating significant coherence in the target–competitor phase difference. For all analyses, n = 21 participants were included.

The next set of analyses was not pre-registered and tested for a relationship between target–competitor phase distance and behaviour. The sample was split according to each individual’s intrusion score in the third repetition, dividing participants into a low-intrusion and a high-intrusion group (Figure 4b). We found that the high-intrusion group had a mean separation angle of 7° that was not significantly different from zero (Rayleigh test for non-uniformity, z(9) = 1.77, p=0.17). The low-intrusion group, by contrast, showed a mean phase separation of 57° that significantly clustered around this angle (Rayleigh test for non-uniformity, z(9) = 3.74, p=0.02); however, we found no significant deviation from zero phase shift in this subgroup of participants (Rayleigh test for non-uniformity with a specified mean angle of 0, z(9) = 3.32, p=0.069). A statistical comparison of the phase separation angle in the high- and low-intrusion groups (Wilcoxon signed-rank test in non-circular space, see ‘Materials and methods’) did not indicate a difference between the mean of the high- and low-intrusion groups’ separation angle (Z = 43, p=0.85). Note that each subsample in this comparison only contains 10 participants, and the statistical comparison is thus likely underpowered. In sum, this analysis suggests that only low-intrusion participants exhibit significant phase separation by the end of the task, which can be taken as an indication that more neural differentiation of overlapping memories along the theta cycle relates to more successful behavioural differentiation.

Two more analyses were conducted to corroborate the finding of a phase difference between target and competitor reactivation, complementing the pre-registered analysis reported above. First, we tested for a temporal difference (lag) between the timelines of target and competitor reactivation using a cross-correlation. To do so, we filtered the two fidelity timecourses from each participant at the dominant 3 Hz frequency, in the third repetition of the CC. We then quantified a temporal lag between timecourses by using a sliding window (allowing us to see phase lags that evolve over time) and calculating the cross-correlation for each 330 ms time bin (1/3 of the dominant 3 Hz frequency). Note that this analysis does not specifically take the phase of virtual hippocampal sensors into account and instead simply cross-correlates the entire target and competitor fidelity timecourses derived from all MEG sensors. The analysis revealed a significant cluster starting approximately 1 s after cue onset and lasting until the end of the trial, with a maximum lag of 30° phase angle (Figure 4c). This finding suggests that the reactivation of competitor memories is lagging 30° behind the reactivation of target memories.

The second additional analysis used the average target and competitor fidelity timecourse of each participant, filtered at 3 Hz and shown individually in Figure 4—figure supplement 1a. Visual inspection already suggests a temporal shift, and sometimes even phase opposition, between target and competitor decoding in most participants. To formally quantify the phase shift on a group level, we subtracted in complex space, for each time bin, the phase angle of the target from the angle of the competitor. The resulting phase separation (Figure 4—figure supplement 1b) is consistent with the cross-correlation results shown in Figure 4c and suggests a continuous lag of approximately 45° between the two competing memories. The results of these two analyses on the continuous fidelity timecourses are thus consistent with the phase shift identified when looking at the classifier-peak-to-hippocampal-theta locking, and provide further evidence for a temporal segregation between target and competitor reactivations during the last retrieval repetition.

Discussion

To remember a specific experience, the respective memory often needs to be selected against overlapping, competing memories. The processes by which the brain achieves such prioritisation are still poorly understood. We here tested a number of predictions derived from the oscillating interference resolution model (Norman et al., 2006). In line with its predictions, the present results demonstrate that the neural signatures of two competing memories become increasingly phase separated over time along the theta rhythm. Furthermore, larger phase differences were associated with fewer intrusion errors from the interfering memories. These findings support the existence of a phase-coding mechanism along the cycle of a slow oscillation that adaptively separates competing mnemonic representations, minimising their temporal overlap (Lisman and Idiart, 1995; Norman et al., 2006).

Based on the oscillating interference resolution model (Norman et al., 2006), we hypothesised that target memories would be strengthened and become decodable at an earlier (higher inhibition) phase (see Figure 1d), while competing associates would be weakened, with time windows of their reactivation gradually shifting to a later (low inhibition) phase. The phase separation between the two competing memories should thus increase the more often the competing memories are being reactivated, and we expected to find an observable difference in phase separation after a number of repeated recalls. To further test whether a large phase segregation is beneficial for memory performance, we related phase distance to the number of behavioural intrusion errors. Supporting our hypotheses, in the third and final recall repetition target and competitor memories showed a significant phase shift with an average distance of 34° (Figure 4a and b). Larger phase separation angles in individual participants were related to lower levels of behavioural interference. Participants with low levels of memory intrusions showed a significant phase difference of 57°, while high-intrusion participants showed a non-significant 7° phase shift on average (Figure 4b). Also in the last repetition, a significant phase modulation was found for both target and competitor memories such that the fidelity of their neural reactivation was rhythmically modulated by the dominant 3 Hz rhythm (Figure 3e and f). Finally, a significant cluster of cross-correlation of the continuous target and competitor decoding timecourses also suggested an average phase lag of 30°, complementing the analyses that explicitly relate classification performance to hippocampal theta phase. Though deviating from the pre-registered hypotheses in some respects as discussed below, these results largely confirm our predictions derived from the oscillating interference resolution model (Norman et al., 2006), supporting a central role for slow oscillations in orchestrating memory recall by temporally segregating potentially interfering memories.

Based on the same model (Norman et al., 2006), we also predicted that across repeated retrieval attempts, target memories would be strengthened and competitors become less intrusive. Though not tested behaviourally in this study, repeated recall is known to induce enhancement of the retrieved memories (Karpicke and Roediger, 2008; Rowland, 2014) and forgetting of non-retrieved, competing memories (Anderson et al., 1994). Evidence for the up- and downregulation of neural target and competitor patterns, respectively, has previously been shown in fMRI studies (e.g. Kuhl et al., 2007; Wimber et al., 2015). In this work, we found neural evidence for both target strengthening and competitor weakening within the time window of maximum memory reactivation (Figure 2h). The increase in target evidence was more robust than the decrease in competitor evidence, with the latter not surviving multiple comparisons correction across the entire recall time window. It is possible that competitor decoding across all trials was too noisy in the first place, especially when focusing on trials where participants correctly answered both follow-up questions, and where neural patterns should thus be dominated by retrieval of the target memory. When instead analysing only incorrect trials, we found moderate evidence for the neural reactivation of competitor memories (Figure 2—figure supplement 2d). The relatively stable levels of competitor reactivation over repetitions might also be due to the lack of feedback in our task, allowing little adjustment of incorrect responses, as also reflected in the flat behavioural intrusion index (Figure 2b, Figure 2—figure supplement 1a). Overall, however, and in line with recent EEG work (Bramão et al., 2022), our results suggest that dynamic changes in overlapping memories can be tracked using MEG.

Decoding of target and competitor memories in the high-interference (CC) condition was generally less robust than decoding of target memories in the low-interference (NC) conditions. Generally, one challenge for time-resolved decoding of reactivated memory content is the considerable variance in the timing of memory recall across trials, conditions, and participants, likely affecting the timing in neural pattern reinstatement. Such variance can be rectified, at least to a degree, when aligning the timelines to the button press that indicates subjective recall, as shown by our response-locked analyses leading to more robust target decoding (see Figure 2f and g). There are, however, several reasons why we pre-registered our analysis locked to the onset of the memory cue. Most importantly, previous work suggests that the phase of theta oscillations is reset by a memory cue and remains relatively stationary for a period of time (see Rizzuto et al., 2006; Ter Wal et al., 2021), such that the locking of memory reactivations to the theta rhythm should be most coherent when aligned to cue onset. Note that all crucial analyses comparing phase shifts of target and competitor reactivations were conducted on the level of single trials. Therefore, these analyses are not affected by whether or not the reactivation maxima are consistent in time across trials and participants, even though such inconsistencies will dilute the average cue-locked decoding performance.

The functional relevance of slow oscillations in regulating the excitation–inhibition balance of specific neural assemblies has featured strongly in other computational models too, and these models are supported by our findings at various levels. Most relevant in the present context is the theta-gamma model (Lisman and Idiart, 1995; Lisman and Jensen, 2013), which assumes that the coupling of gamma bursts to specific theta phases provides a timing mechanism for ordering sequences of items in working memory. Here, a gamma burst reflects the firing of a local cell assembly representing a distinct unit of information, while the phase of the slower theta oscillation rhythmically modulates the excitability of these local neural assemblies, and feedback inhibition allows for an organised, non-overlapping firing sequence across the theta cycle. In humans, studies using episodic and working memory tasks have provided evidence for such a theta-gamma code (Bahramisharif et al., 2018; Canolty et al., 2006; Fuentemilla et al., 2010; Griffiths et al., 2021; Herweg et al., 2020; Heusser et al., 2016; Karlsson et al., 2022; Reddy et al., 2021; Ten Oever et al., 2020; Ter Wal et al., 2021). The oscillating interference resolution model (Norman et al., 2006) and theta-gamma phase-coding models share the fundamental assumption that slow oscillations regulate the excitation–inhibition balance of local neural assemblies (Buzsáki, 2006). However, they do differ in their theoretical scopes. Theta-gamma code models were developed to explain how the brain can handle ongoing states of high attentional or working memory load, where multiple distinct items need to be kept separated and organised. The oscillating interference resolution model (Norman et al., 2006) is a learning model and thus explicitly oriented towards optimising future network states. It employs fluctuating inhibition to actively shape memory representations, with the goal to minimise future overlap and interference.

To our knowledge, no study in humans has explicitly tested for phase coding when several overlapping memories compete for retrieval. However, one intracranial study using a virtual navigation paradigm showed that the neural representations of potentially interfering spatial goal locations are coupled to distinct phases of the hippocampal theta rhythm (Kunz et al., 2019). Phase precession and theta coupling of neuronal firing during goal-directed navigation have also been shown in human hippocampal single-unit recordings (Qasim et al., 2021; Watrous et al., 2018). The present results directly support the idea that phase coding can temporally separate co-active, competing memory representations in the human brain. They additionally suggest that such a code adaptively evolves across repeated target retrievals, in line with the idea that oscillating excitation–inhibition can optimise learning with respect to future access to the relevant target memory.

A second prominent computational model focuses on different functional operations within the hippocampal circuit at opposing phases of the theta rhythm, with one half of the hippocampal theta cycle devoted to the encoding of new incoming information, and the other half to the recovery of internally stored information (Hasselmo et al., 2002; Kunec et al., 2005). In humans, previous EEG work Kerrén et al., 2018 and a recent study investigating rhythmic modulations of reaction times Ter Wal et al., 2021 have provided evidence for such process separation. In the present study, target and competitor reactivations peaked at distinct theta phases, with an average separation angle of approximately 35°, and 57° in the low-intrusion group that presumably reached optimal interference resolution. This observation is consistent with a framework where both targets and competitors are reactivated during the retrieval portion of a theta cycle (Hasselmo et al., 2002; Kunec et al., 2005), but with sufficient phase separation to allow for differentiation in time. Note, however, that the oscillating interference resolution model (Norman et al., 2006) dedicates the entire theta cycle to retrieval operations (as depicted in Figure 1d) and would thus predict temporal distances of up to 180° between targets and competitors. The two models may not necessarily be incompatible, especially when considering that encoding and retrieval computations take place in different subcircuits of the hippocampus (Hasselmo et al., 2002). Future modelling work is needed, however, to better understand the learning dynamics across the full theta cycle and reconcile the apparent theoretical conflict between the two frameworks.

Other predictions of the oscillating interference resolution model (Norman et al., 2006) relate to the neural overlap in spatial patterns and have been tested primarily in fMRI studies. One such prediction is the non-monotonic plasticity hypothesis, derived from the same learning dynamics along the oscillatory cycle as described above. If a competitor is sufficiently strong to be co-activated in the high-inhibition (target) phase, it will benefit from synaptic strengthening of common connections and thus be integrated with the target memory. Moderately co-active competitors, on the other hand, will be subject to synaptic depression and as a result become differentiated from the target memory (Ritvo et al., 2019). Evidence for non-monotonic plasticity comes from fMRI work that directly tracks changes in the similarity of target and competitor representations in terms of their spatial patterns (Hulbert and Norman, 2015; Wammes et al., 2022), and the representational changes dependent on the level of (co)activation of overlapping memories (Detre et al., 2013; Lewis-Peacock and Norman, 2014; Poppenk and Norman, 2014; Wang et al., 2019). Indirectly related, fMRI studies investigating representational change resulting from repeated learning or recall have also found evidence for differentiation or integration of overlapping memories in hippocampus, late sensory cortex, and prefrontal cortex (Favila et al., 2016; Hulbert and Norman, 2015; Kim et al., 2017; Kuhl et al., 2010; Schlichting and Preston, 2015; Wimber et al., 2015), and sometimes observed both in the same study but in different regions (Molitor et al., 2021; Schlichting and Preston, 2015). The present study was not designed to track representational changes, for example, in terms of comparing target and competitor representations before and after repeated recall. Arguably, however, there is a straightforward agreement between spatial and temporal separation in this network model. Strong competitors would tend to co-activate in the high-inhibition phase and thus be co-strengthened and integrated via LTP (Bliss and Lomo, 1973; Hebb, 1949), whereas moderately co-active competitor memories would tend to co-activate only at lower levels of inhibition, which according to the model would lead to LTD. Our data support the idea that low temporal separation is related to integration, behaviourally. The high-intrusion group had a mean phase difference of 7° between target and competitor memories, theoretically in line with a time window of strong co-firing hence synaptic strengthening. Integration might thus have led to higher amounts of intrusions (Brunec et al., 2020). In the low-intrusion group, more pronounced phase separation between target and competitor memories (57° on average) might have led to only moderate co-firing and hence promoted differentiation, resulting in lower levels of behavioural interference. Although speculative, temporal separation could be a prerequisite for spatial differentiation. Future studies, combining high temporal with high spatial resolution, and paradigms to track the representational distance of overlapping memories, are needed to fully understand these dynamics.

A priori we included the entire theta band in our pre-registration as previous studies have indicated large variations in the peak theta frequency in humans (Goyal et al., 2020; Jacobs, 2014; Kerrén et al., 2018; Lega et al., 2016; Ter Wal et al., 2021). For example, in our previous study we found that reactivations of target memories were most consistently coupled to a hippocampal 7–8 Hz oscillation, even though a slower 3–4 Hz peak was also present (Kerrén et al., 2018). In another study, Ter Wal et al., 2021 found that slower 2–3 Hz oscillations dominated behaviour and hippocampal signals specifically when responses depended on memory, consistent with other intracranial work reporting pronounced slow theta oscillations in the human hippocampus during memory tasks (Goyal et al., 2020; Lega et al., 2012). In working memory, it has even been demonstrated that theta oscillations slow down with increasing memory load (Fell et al., 2011; Wolinski et al., 2018), in line with the theoretical idea that slower rhythms accommodate better phase separation (Lisman and Idiart, 1995; Lisman and Jensen, 2013). Descriptively, such theta slowing was evident in the present data when comparing the frequency profiles of rhythmic memory reactivation between NC and CC conditions, where conditions in which only one associate needs to be remembered showed slightly faster rhythmic peaks (4–5 Hz) than the condition in which two associates compete for retrieval (3–4 Hz, see Figure 3—figure supplement 1). While offering a possible explanation why slower oscillations are often observed in tasks with high memory demands, this post-hoc interpretation will need to be corroborated by empirical evidence in future studies.

Phase coding offers an elegant theoretical solution to temporally segregate the processing of potentially interfering information. Much empirical evidence for phase sequencing along the hippocampal theta oscillation comes from spatial navigation work, both in rodents and humans (Herweg et al., 2020; Kunz et al., 2019; O’Keefe and Recce, 1993; Reddy et al., 2021). We here demonstrated that phase coding facilitates the separation of overlapping, associatively linked memories. Together with the computational model used to derive these predictions, these findings offer a possible mechanism utilised by the human brain to resolve competition between simultaneously active memories. More generally, they add to a growing literature showing that slow oscillations orchestrate the intricate timing of neural processing with direct, observable effects on behaviour.

Materials and methods

Subjects

Twenty-six right-handed participants (18 females, 8 males) took part for financial or course credit compensation (mean = 24.1 years, SD = 5.73 years, range 18–33). They all had normal or corrected-to-normal vision and reported no history of neurological disorders. All experimental procedures in this study were approved by and conducted in accordance with the University of Birmingham’s STEM Research Ethics Committee (ERN-16-1512 and ERN-18-0226P). Written informed consent was obtained from participants before they took part in the experiment.

Stimuli and task

Request a detailed protocol

The material consisted of 72 images depicting animate and inanimate objects (equal number of mammals, birds, insects, and marine animals, electronic devices, clothes, fruits, and vegetables), and 72 images depicting indoor and outdoor scenes, taken from the BOSS database (Brodeur et al., 2010) and online royalty-free databases. Stimulus selection was motivated by previous success at distinguishing these categories using time-resolved multivariate pattern analysis (Kerrén et al., 2018). Images from both object and scene classes were pseudo-randomly split for each participant into six sets, so that each set consisted of 20 objects and 20 scenes, 10 animate and 10 inanimate, 10 indoor and 10 outdoor scenes. Each set constituted one learning block. Also, a list of 144 action verbs, largely overlapping with those used in Linde-Domingo et al., 2019, served as cue words in the cued recall task. These words were randomly assigned to images and conditions for each participant for creating the relevant word–image associations. Three additional associations were used for demonstrative purposes.

Participants received task instructions and first performed one short practice block. All participants (see exclusion below) then performed six experimental blocks (40 encoding trials and 72 (24 × 3) retrieval trials per block), each consisting of an associative learning phase, a distractor task, and a retrieval test with three repetitions per target item (Figure 1). At encoding, participants were asked on each trial to encode a word together with an image associate. In the CC, a word was encoded together with two associates, separated by at least three intervening trials. The instruction was to always memorise the most recent associate that was presented together with a given word for the subsequent memory test. Therefore, the second associate in the CC always served as the target, with the previously learned first associate (i.e. competitor) assumed to elicit proactive interference. In the non-competitive single-exposure condition (NC1), a word was encoded together with only one associate, and these associations were never repeated during encoding. This condition served as the behavioural baseline for measuring the effect of proactive interference on memory performance (i.e. having previously encoded a competing associate compared with only one associate). In the non-competitive double-exposure condition (NC2), participants also encoded a word together with only one associate, but these associations were presented twice. This condition served as the neurophysiological baseline and was specifically designed to control for neural effects induced by the repetition of the word cue (including but not limited to repetition suppression; Epstein et al., 2008; Kristjánsson and Campana, 2010). The conditions were shown in pseudo-randomised order within each block, where 1/3 of the associations were in NC1, 1/3 in NC2, and 1/3 in CC (NC1 = 8 associations, NC2 = 8 associations, CC = 8 associations). In NC1 and NC2, there were an equal number of objects and scenes (four objects and four scenes per block, but shown twice each for NC2, hence resulting in eight object and eight scene trials), whereas in CC there were eight object and eight scene trials, with one pair each linked to the same cue word. In total, this summed up to 40 trials for one block of learning (NC1 = 8 trials, NC2 = 16 trials, CC = 16 trials). The order of the trials belonging to the three conditions was pseudo-randomised such that the average serial position of each condition within a block was equal. Images were pseudo-randomly assigned to these conditions for each participant, with the constraint that in the CC the associates needed to be from different image categories (one object and one scene, split such that on half of the CC trials, the target was an object and the competitor a scene image, and vice versa for the remaining half).

A learning trial consisted of a jittered fixation cross (between 500 and 1500 ms), a unique action verb (1500 ms), a fixation cross (1000 ms), followed by a picture of an object or scene that was presented in the centre of the screen for 4 s. Participants were asked to come up with a vivid mental image that linked the image and the word presented in the current trial. As soon as they had a clear association in mind, they pressed the right-thumb key on the button box. Participants were aware of the later memory test.

A distractor task followed each learning phase. Here participants had to indicate whether a given random number (between 1 and 99) presented on the screen was odd or even. They were instructed to accomplish as many trials as they could in 45 s and received feedback about their accuracy at the end of each distractor block.

After the distractor task, participants’ memory for the 40 verb–object associations memorised in the immediately preceding learning phase was tested in pseudo-random order, with three repetitions per relevant association. The order was pseudo-randomised such that all associations were first tested once in random order before testing all associations again in a new random order, and again a third time. No significant difference in serial position resulted when comparing CC and NC2 items [t(23) = –1.69, p=0.1], CC and NC1 items [t(23) = –1.18, p=0.25], or NC1 and NC2 items [t(23) = –0.43, p=0.66]. One repetition consisted of 24 trials (eight for NC1, eight for NC2, and eight for CC), with the constraint that a given association could only be tested once per repetition. Each trial consisted of a jittered fixation cross (500–1500 ms), followed by one of the words as a reminder for the association. Participants were asked to bring back to mind the most recent associate of this word as vividly as possible. The cue was presented on the screen for 500 ms and thereafter a blank screen with a black empty frame was presented. To capture the particular moment when participants consciously recalled a specific association, they were asked to press the right-thumb key as soon as they had a vivid image of the associated memory in mind. If they did, the frame flashed once, and participants were presented for 4 s with a blank screen and asked to hold the image in mind. A question then appeared on the screen asking whether the retrieved item was an object, a scene, or they were unable to remember. Across trials, the object and scene options randomly shifted between the left and right sides of the screen. If the participant did not remember the association, they were told to press the left-thumb button. If participants selected ‘object’ or ‘scene,’ a follow-up question appeared (dependent on the response to the first question), asking whether the retrieved associate was an inanimate or animate object, or whether it was an indoor or outdoor scene. The two follow-up questions were self-paced, and there was no feedback.

The experiment was set up via custom-written code in MATLAB 2016a (The MathWorks, Munich, Germany) using functions from the Psychophysics Toolbox Version 3 (Brainard, 1997). The presentation was projected onto a screen located 1.5 m away from the participant using Windows 64 bit. Images were 2000 × 2000 pixels large.

Analysis

The general hypotheses and analysis steps were pre-registered and can be found on OSF (here).

After excluding participants based on the criteria stated in the pre-registration (more than 2 SD from the mean accuracy for each condition separately), 24 participants (16 females, 8 males) remained, with an average age of 24.5 years (SD = 5.73). For the MEG analysis, a further 3 participants were excluded due to noisy data, which resulted in 21 participants included in the analyses. All participants performed all six blocks except for two, for whom time limit and button box errors occurred, resulting in only five blocks for these participants for MEG analysis.

Statistical analysis

Request a detailed protocol

In view of the pre-registration on OSF, all statistics were one-sided and conducted on a group level. Non-parametric cluster-based permutation tests were conducted to test classification performance against chance across multiple time points, and for time–frequency analyses comparing CC and non-CC (see below). Only correct trials were used for all MEG analyses. Where a specific frequency band had to be preselected, the analyses were limited to 3–8 Hz. This, however, deviated from the frequency range of interest in the pre-registration, which was 4–8 Hz. Since the study was pre-registered, several studies have shown slower theta oscillation specifically related to human episodic memory (Goyal et al., 2020; Ter Wal et al., 2021) and we therefore decided to include also 3 Hz in the frequency band of interest.

MEG data analysis

Request a detailed protocol

The MEG was recorded at the Centre for Human Brain Health (CHBH), Birmingham, UK, using an Elekta Neuromag TRIUX system, with 306 channels (204 planar gradiometers and 102 magnetometers; only gradiometers are used for all analyses reported here), sampled at 1000 Hz (Elekta, Stockholm, Sweden). EEG was recorded with a 64-channel electrode cap in the initial 10 participants as a sanity check in order to verify that a strong mid-frontal theta signal can be observed in the CC > NC2 condition, and how its topography compares between MEG and EEG (Cavanagh and Frank, 2014). Since we found a comparably strong theta increase over EEG and MEG (gradiometer) sensors, the EEG data is not reported in any of the analyses presented here. The experiment was shown on a projector screen using a PROPixx projector (VPixx Technologies, Saint-Bruno, Canada) with a 1440 Hz refresh rate, and participants’ responses were collected using two button response boxes (fMRI Button Pad [2-Hand] System, NAtA Technologies, Coquitlam, Canada).

For source localisation, individual anatomical MRI scans (T1-weighted; 1 × 1 × 1 mm voxels; TR = 7.4 ms; TE = 3.5 ms; flip angle = 7°, field of view = 256 × 256 × 176 mm) were acquired at the CHBH (3T Achieva scanner; Philips, Eindhoven, the Netherlands).

Preprocessing

Request a detailed protocol

Preprocessing was done using the FieldTrip toolbox versions 2019 and 2021 (Oostenveld et al., 2011) and custom-written MATLAB code. A Butterworth low-pass filter of 200 Hz and a band-stop filter (50 Hz, 100 Hz, and 150 Hz) were applied to the data. To avoid temporal displacement of information, while still removing slow drifts, trial-masked robust detrending was applied to the retrieval data, in accordance with recent recommendations (van Driel et al., 2021). Briefly, the data for this correction step were divided into epochs spanning 15 s before cue onset and 15 s after cue onset. The experimentally relevant events started 1000 ms before cue onset until 4000 ms after cue onset. This time window was masked out from each trial. To make sure all data would be included, the continuous data were symmetrically mirror-padded with 15 s prior to segmentation. To improve the fit of the higher-order polynomial, a first-order polynomial was used to detrend the entire epoch (in accordance with de Cheveigné and Arzounian, 2018). Thereafter, a 30th-order polynomial was fitted and removed from the data. Note that the events of interest were not a part of the fitting procedure, making sure the fit was not being influenced by cognitively relevant processing. The method was partly implemented using the Noise Tools toolbox (http://audition.ens.fr/adc/NoiseTools) together with custom-written MATLAB code. After this step, the detrended data were cut into the experimentally relevant epochs (−1000 to 4000 ms around cue onset at retrieval).

An automatised trial and component rejection was applied to the data. In a first step, to remove high-frequency bursts, data were high-pass filtered at 100 Hz, and trials that exceeded four times the median absolute deviation of the amplitude distribution across trials were automatically removed. In a second step, independent component analysis (ICA) was used to detect artefacts to be removed in the data. To this end, MEG data were downsampled to 250 Hz, and only the first 1.5 s after cue onset were used for ICA (to reduce computational load). The unmixing matrix was then applied to the entire epoch of non-downsampled data. To remove blink and cardiac artefacts, a template was created by running an ICA on the six first participants. The components for the two artefacts were then averaged separately, and each average component was used as a template to identify matching components in each participant. For the blink template, the algorithm correlated its topography with that of each component identified in an individual participant, and subsequently removed the component with the best match. For the cardiac component, which is difficult to identify based on topography, the algorithm filtered the component time series to between 10 and 20 Hz and correlated the peak amplitude of each trial with the cardiac time-series template. Any component which correlated to a greater extent than four times the median absolute deviation of the correlation was classed as a cardiac component and removed. All removed components were visually inspected for validation.

After these components were removed, an additional automatised artefact rejection was conducted, similar to the initial step. The same procedure as in the first rejection round was followed but was done on channels instead of trials. Again, data were band-pass filtered between 1 and 100 Hz, and the channels that exceeded three times the median absolute deviation of the channel distribution were rejected. We chose three times the median absolute deviation for the last two rounds of rejection to be more conservative. Bad channels were interpolated using the triangulation method (Oostenveld et al., 2011). After components and trials were removed, all trials and components were again visually inspected, and trials and components still containing artefacts were manually removed. Lastly, data were downsampled to 250 Hz. Due to poor data quality in 3 participants, remaining data from 21 participants were used for MEG analyses. On average, 353 out of 432 retrieval trials were kept (min = 290, max = 397, SD = 27) for analysis. In the CC, the number of trials on average was 118 (max = 140, min = 95, SD = 10.24). Per retrieval repetition there was an average of 40 (SD = 3.81), 39 (SD = 3.63), and 38 (SD = 4.31) trials for each respective repetition. For the NC1 and NC2 collapsed, the average number of trials per retrieval repetition were 79 (SD = 7.38), 77 (SD = 6.36), and 77 (SD = 7.29). At the end of preprocessing, a sanity ERF-check on occipital channels of the retrieval data was conducted and the average waveform can be found in Figure 3—figure supplement 2a.

Time–frequency decomposition

Request a detailed protocol

The spectral difference between the CC and neural baseline (NC2) during retrieval was calculated by convolving the combined activity of each planar channel (summing the magnitude of the planar gradient over both directions at each sensor) with a complex Morlet wavelet of a minimum of five cycles (increasing with frequency to cover approximately 500 ms in length [e.g. 10 Hz: 10 * 0.5 = 5 cycles]) from 1 to 20 Hz for each condition. First, paired-samples t-tests were computed between the two conditions to investigate the difference in spectral power. To account for multiple comparisons across time points (0–1000 ms post-cue), frequencies (3–8 Hz), and sensor (102 combined gradiometers), the t-statistics were subjected to non-parametric cluster-based permutation testing, as implemented in the FieldTrip software. The threshold for statistical testing was set to a cluster alpha level of 0.05. The minimum number of neighbouring channels that were considered a cluster was set to three. t-Values above the threshold of 0.1 were then summed up in a cluster and compared against a distribution where condition labels were randomly assigned 1000 times with the Monte Carlo method, following the default method implemented in FieldTrip.

Multivariate pattern analysis

Request a detailed protocol

Two separate LDA-based classifiers were trained on discriminating subordinate picture class, that is, animate vs. inanimate for objects, and indoor vs. outdoor for scenes. This important feature of the design allowed us to have two classifiers that provide independent measures of target and competitor reactivation. The two classifiers were trained on the non-CCs (‘pure’) (i.e. retrieval of objects and scenes in the NC1 and NC2 conditions), and were then tested on both the non-CCs and CCs. The time interval of interest for these multivariate analyses started 500 ms before onset of the retrieval cue and lasted up to 3000 ms post-cue. This time window was selected because no memory reactivation is expected earlier than 500 ms (Staresina and Wimber, 2019), and previous work using a similar (though non-competitive) cued recall paradigm suggests that participants need approximately 3 s to mentally reinstate a image (Linde-Domingo et al., 2019; Kerrén et al., 2018). A sliding Gaussian window with a FWHM in the time domain of 40 ms was applied to the trial time series before classification analysis. Each retrieval trial was then baseline corrected by subtracting the mean amplitude in a time window from 400 to 50 ms pre-cue, separately per sensor. Training and testing were conducted time point per time point, in steps of 8 ms, each time point centred at the centre of the Gaussian smoothing window. The MEG gradiometer sensor patterns (amplitudes on each of the 204 channels at a given time point) were used as classifier features, independently per participant.

Training/testing in the non-CC was done as a sanity check for classifier performance in a memory retrieval situation with no interference. Since the same data was used for training and testing in this case, a tenfold cross-validation was used and repeated five times. To test for reactivation of target and competitor memories in the CC, trials were split into those where the target was an object and the competitor was a scene, and vice versa, and the corresponding object and scene classifiers (trained on non-competitive retrieval) were then used to separately indicate evidence for target and competitor reactivation on each single trial. Only correct trials were used in the CC. As the training and testing data came from different trials, cross-validation in the CC was not necessary. To avoid overfitting, the covariance matrix was regularised using shrinkage regularisation, with the lambda set to automatic (Blankertz et al., 2011).

The LDA, as used here, reduces the data from 204 channels into a single decoding timecourse per trial, and we used these single-trial, time-resolved decision values (d-values or fidelity values) of the classifier as an index of memory reinstatement (Carlson et al., 2014). Classifier accuracy was derived by calculating the fraction correctly predicted labels, whereas chance was defined as 50% for a binomial classifier. More specifically, during training, the classifier found the decision boundary that could best separate the patterns of activity from the two classes (animate vs. inanimate for objects, indoor vs. outdoor for scenes) in a high-dimensional space. The classifier was then asked to estimate whether the unlabelled pattern of brain activity in any given retrieval trial and at each time point was more similar to one or the other class. This training test procedure was repeated until every single retrieval trial had been classified. A larger smoothing kernel of 200 ms FWHM in the time domain was applied to the decoding timecourses purely for visualisation but not for statistical analysis purposes.

As part of a (not pre-registered) set of analyses conducted in response to peer review, we also analysed decoding performance after realigning the data to the button press that indicates subjective recollection of the target associate. These analyses followed the same approach as above, except that we excluded trials in which responses occurred in the first 500 ms or last 200 ms to be able to plot decoding accuracy from –500 to +200 around the button press.

Determine peak frequency of fidelity values using IRASA

Request a detailed protocol

The IRASA method has been shown to robustly find and separate the oscillatory from the fractal signal both in ECoG and MEG data, and was here used to quantify the oscillatory signal component of the fidelity values (Wen and Liu, 2016). More specifically, the brain produces task-related rhythmic (oscillatory) components, but also arrhythmic scale-free (fractal) components (Buzsáki and Draguhn, 2004). The rhythmic oscillatory components are regular across time, whereas the fractal components are irregular (Wen and Liu, 2016). In short, IRASA resamples a time-series signal and computes a geometric mean of every pair (oscillatory and fractal) of the resampled signal. The median of the geometric mean is then used to extract the fractal power spectrum. The difference between the original power spectrum and the fractal power spectrum is the estimate of the power spectrum of the oscillatory component of the signal (Wen and Liu, 2016). In this study, IRASA (as implemented in the FieldTrip toolbox) was applied to the fidelity values, in a time window from 500 to 3000 ms post-cue, padding each trial length up to the next power of 2. Apart from the reasons mentioned above, this large time window also assures that low frequencies can be properly estimated.

To test for significance in the frequency-transformed decoding timecourses, the LDA as described above was repeated 15 times per subject, now trained on random labels in each iteration. The fidelity values from each iteration were then subjected to the same IRASA method as the non-shuffled data. An empirical null distribution was created on the second (i.e. participant) level as follows. In each of 1000 repetitions, one of the 15 surrogate-label classifier outputs (i.e. the oscillatory component output by IRASA) per participant was randomly selected, and an empirical group average was computed for this random selection. This procedure resulted in 1000 group averages of oscillatory strength estimates for each given frequency of interest, representing the chance distribution at this frequency. The IRASA output of the real-label classifier could then be contrasted against this chance distribution (Cohen, 2014). Note that due to the algorithm for separating the fractal components from the oscillatory components, the output from IRASA yields a higher frequency resolution than 1 point per frequency. However, our a priori frequency range was set to 3–8 Hz, and we therefore tested for significant oscillatory components in each 1 Hz frequency bin between 3 and 8 Hz, with the estimated chance distribution subtracted from the real value and subsequently divided by the standard deviation of the estimated chance distribution (Figure 3a and b). This gave a z-value, which was compared to the critical threshold of z = 2.32 at p=0.01, correcting for five multiple comparisons (3–4 Hz, 4–5 Hz, 5–6 Hz, 6–7 Hz, and 7–8 Hz). Again, this procedure replicated the one used in Kerrén et al., 2018.

Source analysis

Request a detailed protocol

The aim was to extract activity from hippocampal virtual channel, hence, the magnetometers were used and their data projected into source space using an LCMV beamforming approach (Gross et al., 2007). Individual MRI scans (anatomical data) were available for 18 participants, and for the remaining 3 participants the standard head model as implemented in FieldTrip was used. Anatomical scans were aligned based on the sensor position obtained from a Polhemus system (Colchester, Vermont, USA). This was done in three steps: the first step was done manually by adjusting the alignment between the anatomical data and the head position; the second was done by the Iterative Closest Point algorithm implemented in the FieldTrip toolbox Oostenveld et al., 2011; and the third step was again done manually to check that the alignment worked, and adjusted if it did not. The realigned model was used to reslice and segment the brain to make the axes of the voxels consistent with the head position and subsequently to extract the brain compartments. The segmented brain was then used to create a forward model (head model). In our case, we used a semi-realistic forward model (Nolte, 2003). The forward model was used to create the source model (lead field), where for each grid point the source model matrix was calculated with a 1 cm resolution, and the virtual sensors were placed 10 mm below the cortical surface, and subsequently warped into each brain. In total, we modelled 3294 virtual sensors for each participant with whole-brain coverage. LCMV beamforming was used to reconstruct the activity of all virtual channels in source space (see below for selecting hippocampal channels) with 20% regularisation. To confirm that the source localisation provided reliable results, we checked sources for the early visually elicited response to the cue word in an early time window by calculating the pre-/post-variance around cue onset (–200 ms to 0 and 0 to 200 ms). Because of arbitrary orientation of the dipole in source space, variance is an unbiased measure to investigate where in space and when in time activity fluctuates (see Figure 3—figure supplement 2b).

For the phase-amplitude coupling between source-localised hippocampal 3 Hz phase and fidelity values, the filters were computed on 4 Hz low-pass-filtered data, from –1000 ms to 3000 ms after cue onset (baseline-corrected [–400 to –50]). The filters were obtained from correct trials in the competitive condition. The left and right hippocampus were chosen as a region of interest, derived from the AAL atlas (Oostenveld et al., 2011). The individual MRI and the atlas were interpolated using the interpolation method nearest, finding the nearest region in the AAL atlas based on the Euclidean distance. Single-trial timecourses were extracted from the corresponding virtual sensors, and activity averaged across those sensors in the left and right hippocampus, resulting in one activity timecourse per trial for left and right hippocampus, collapsed.

Further source-level analyses were conducted to estimate the frequency profile of the raw trial data in the hippocampal region of interest and compare it to control regions (superior occipital cortex and precentral gyrus), with the hypothesis that the hippocampus would show a stronger power in the theta frequency range (3–8 Hz) than other regions. We used IRASA on each trial and each virtual channel, following the same procedure as we did when calculating the frequency profile of the fidelity values. For statistical comparison of the hippocampal profile against each of the two control regions, we directly contrasted the average power in our predefined 3–8 Hz frequency window of interest using two paired-samples t-test. Tests were conducted one-tailed because we expected higher theta power in the hippocampus compared to each of the other two regions. Since we conducted two separate tests (one per control region), we thus set the Bonferroni-corrected p-threshold for each test to.05.

Phase-amplitude coupling between MEG data and fidelity values

Request a detailed protocol

The MI (Tort et al., 2010) was calculated to test for a relationship between the fidelity values and a hippocampal theta oscillation, with 3 Hz selected as the frequency of interest based on the results from previous steps (IRASA, see above). The signal from virtual channels was transformed to an analytic signal by convolving the raw data with a complex Morlet wavelet with cycles increasing with frequency to be approximately 500 ms in length for each frequency as implemented in FieldTrip (Oostenveld et al., 2011). Each complex value data point was then point-wise divided by its magnitude, resulting in a 4D matrix of phase values, containing trials * channels * frequencies * time. Phase values at 3 Hz were divided into 10 adjacent bins, ranging from – pi to pi, following the pre-registered procedure. To link classifier-based memory reinstatement indices to the hippocampal phase, the amplitudes of the single-trial fidelity values from corresponding time points were sorted into their respective phase bins, and the average classifier amplitude of each phase bin was calculated.

Following this binning procedure, two metrics were of relevance: (1) the MI showing the extent to which the fidelity values are modulated by the hippocampal phase, and (2) the maximum bin for target and competitor reinstatement. The MI was obtained by creating a uniform distribution and calculating the Kullback–Leibler distance and subsequently the MI, in accordance with the approach originally proposed by Tort et al., 2010. To calculate the pairwise phase difference between the target and competitor reinstatement, the absolute value of the complex number (the real and the imaginary part) representing the average maximum reinstatement peak for each participant was obtained by taking the square root of the sum of the squares of the parts (using the Pythagorean theorem). The amplitude was arbitrarily set to 1 because we were only interested in phase angle. The phase vectors were then point-wise divided with the complex modulus. Finally, to subtract the angles from each other, one vector was rotated by multiplying it with the other’s complex conjugate. To obtain the angle, the inverse tangent of the ratio was taken of the product between the vectors. In MATLAB, the following code was used:

θ=angle(exp1i×X.÷exp1i×Y)

where θ is the vector of angle difference.

We used the CircStat toolbox (Berens, 2009), as implemented in MATLAB, to statistically test the phase difference between target and competitor memories. More specifically, to test the difference in each repetition, as well as the average distance collapsed across all repetitions, we used a Rayleigh test for non-uniformity (circ_rtest function in the toolbox). This test will show significant clustering if the phase differences between target and competitor memories are non-uniform, independent of their absolute mean angle. To test whether the mean phase difference we obtained was significantly different from zero phase shift, we used the circ_vtest function with a specified mean angle of 0. This test indicates whether the mean angle of the phase distance is significantly different from zero. Lastly, we wanted to test the difference between the high- and low-intrusion groups’ mean angle. To do so, we calculated the mean phase angle of target and competitor reactivation for each participant. We then calculated the circular mean of the two vectors within each participant, where the length of this vector defines the similarity between target and competitor phase angles, independent of absolute angle. This vector is non-circular and can be subjected to a Wilcoxon signed-rank test. We chose a Wilcoxon signed-rank test as the sample size was low in each subgroup (n = 10 in each group, removing one participant from the larger group randomly to equate group size).

Cross-correlation and maximum phase difference between fidelity timecourses of target and competitor

Request a detailed protocol

As another part of the set of new analyses conducted in response to peer review, we wanted to establish that the phase shift between target and competitor reactivation can also be observed when using the continuous fidelity timecourses, rather than zooming in on the reactivation peaks only. Single-trial fidelity timecourses from the third recall repetition (correct trials only), as used for the above phase-to-classifier locking analyses, were filtered at 3 Hz, and their lag was quantified using a cross-correlation. A sliding-window approach was used to preserve some degree of time information and investigate whether phase lag between signals evolves over time (within trials). Each sliding window had a length of 330 ms (1 divided by 3 Hz), allowing us to express the lag in phase angles from –180 to 180°. For each participant, 25 surrogates were computed where in each iteration, decoding was repeated 25 times, trained on random labels. The same cross-correlation procedure was then conducted on these surrogate classifiers. To obtain a z-value, the surrogate data was subtracted from the real data and subsequently divided by the standard deviation of the surrogate data. Non-parametric cluster-based permutation testing, as implemented in the FieldTrip software, was used to account for multiple comparisons across time points (500–3000 ms) and phase angle (lag: –180 to 180°). The threshold for statistical testing was set to a cluster alpha level of 0.05. t-Values above the threshold of 0.05 were then summed up within a cluster and compared against a distribution where condition labels were randomly assigned 1000 times with the Monte Carlo method, following the default method implemented in FieldTrip.

To further corroborate a phase shift between the average target and competitor fidelity timecourses, we filtered the single trials from the third repetition at the relevant 3 Hz frequency. We then obtained the phase using a Hilbert transform. To statistically quantify the phase shift on a group level, we subtracted, for each time bin, the phase angle of the target from the angle of the competitor (in complex space) for each participant’s averaged fidelity value using the same method as described in the section ‘Phase-amplitude coupling between MEG data and fidelity values’.

Data availability

All data and code used in this project are made available on OSF and Zenodo.

The following data sets were generated
    1. Kerrén C
    2. van Bree S
    3. Griffiths BJ
    4. Wimber M
    (2022) Zenodo
    Phase separation of competing memories along the human hippocampal theta rhythm.
    https://doi.org/10.5281/zenodo.6602054
    1. Kerrén C
    2. van Bree S
    3. Griffiths BJ
    4. Wimber M
    (2022) Open Science Framework
    ID 10.17605/OSF.IO/EVGAW. Target and competitor reactivation in episodic memory.

References

  1. Book
    1. Hebb DO
    (1949)
    The Organization of Behavior; a Neuropsychological Theory
    Wiley.

Decision letter

  1. Huan Luo
    Reviewing Editor; Peking University, China
  2. Laura L Colgin
    Senior Editor; University of Texas at Austin, United States
  3. Ehren L Newman
    Reviewer; Indiana University, United States

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Phase separation of competing memories along the human hippocampal theta rhythm." for consideration by eLife. Your article has been reviewed by 3 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Laura Colgin as the Senior Editor. The following individual involved in the review of your submission has agreed to reveal their identity: Ehren L Newman (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

All the 3 reviewers are impressed by the preregistration study, which makes the hypothesis-testing clear and well-grounded. They also appreciate the large consistency between the results and the tested hypotheses and concur that the findings would be of broad interest to memory and learning fields. Meanwhile, they have also raised several major concerns about the results (see major points below and specific comments by each reviewer).

Essential revisions:

(1) Concerns were raised about the lack of significant memory reactivations for the CC condition, even at the 3rd repetition when memories are supposed to increasingly be reactivated. The authors collapsed target and competitors (Supp Figure 3B) aiming to support memory reactivations, but the rationale is unconvincing since target and competitors are claimed to occur at different phases and should not be combined for decoding.

(2) Related to the 1st point, by comparing reactivations of the 3rd and 1st repetition, the authors suggest a reactivation increase over repetitions. The authors could also consider comparing the 3rd repeat and chance level to examine the memory reactivation at the 3rd repetition. Moreover, how do they reconcile the different reactivation latencies for CC and NC conditions?

(3) Regarding the theta rhythm in classifier fidelity and its relationship to midline theta-band signal, the current results support the phase separation only at the 3rd repetition. It would be important to see how the rhythm of classifier fidelity changes over repeats as well as its characteristics for the NC condition. Moreover, is there any way to confirm that the theta-band rhythm emerges from the hippocampus?

(4) There are other behavioral analyses that could be done to relate the neural findings to the behavior in addition to the intrusion effect (for example: probability of intrusion, RT for memory recall, overall accuracy, etc.). Please see details below in the reviewers' comments.

(5) The authors could consider adding representative data examples to illustrate the phase separation findings.

Reviewer #1 (Recommendations for the authors):

1. The decoding performance didn't reach statistical significance for both target and competitor (Figure 2e), which makes me skeptical to what extent the decoding courses truly denote the memory reactivations. If not, how could the phase-based memory reactivation conclusion be supported? I noticed that the effect mainly occurred at the 3rd repetition recall. Thus, I would recommend that the authors should at least provide evidence that the decoding performance was significant for the target and competitor in the 3rd repetition.

2. The target and competitor showed similar profiles instead of an out-of-phase manner (Figure 1d). Their average decoding performance even showed significant reactivation (supplemental figure 3b). The results seem to be not consistent with the major phase separation findings. Moreover, to flesh out the phase separation results, I would recommend the authors provide typical data examples from their results explicitly illustrating the phase separation profiles, e.g., a representative subject, etc.

3. The key hypothesis is that the neural representation of the target would become stronger, and phase separation would become more prominent across repetitions. They did find stronger reactivation and phase separation for the 3rd repetition compared to the 1st repetition. Meanwhile, is there any corresponding behavioral evidence such that the overlapping memories become less distracting across repetitions?

4. There is no test to confirm the significant theta rhythm in the hippocampus detected in the present study. I think the authors should provide neural evidence backing up that the theta-band rhythm analyzed in their study indeed derives from the hippocampus.

Reviewer #2 (Recommendations for the authors):

I found much to appreciate about this manuscript and work. Much of that is summarized in my 'Public review' so I won't repeat it here.

I had no 'substantial concerns' about the quality of the work or interpretations presented. That is to say, none of the following thoughts that I'll share next should be seen as show-stopping. I will share these thoughts nonetheless in case they can support the growth of this work.

– The reasons why the reactivation effects should only become visible in the 3rd repetition were not clear to me. I struggle to come up with a just-so story based on the Norman et al. 2006 model or otherwise to understand these. Some assistance with this could help.

– There is a strange discrepancy in the frequency band of the frontal midline theta (~7 Hz) and the frequency of the phase modulated reactivations. I did not see any attempt to reconcile this. Are they simply two totally different things, the 7Hz, and 3Hz thetas?

– There are other tests that could be done to relate the reactivation dynamics to the behavior than are described. For example, one could compare the degree of reactivation to the probability of intrusion. Newman et al., 2010 – https://pubmed.ncbi.nlm.nih.gov/20181622/ – for example did something similar. At the coarsest level, the probability of an intrusion should be lowest for trials where there was no evidence of competitor reactivation.

Reviewer #3 (Recommendations for the authors):

There are some suggestions that could improve the overall conclusions one can draw from the manuscript.

The manipulation of repeating the retrieval period is interesting and allows for some novel hypotheses and questions. The authors use a measure of intrusions (by assessing how often subjects select the competitor subordinate category) and find that this is independent of the recall trial number. How about accuracy for the CC conditions?

The authors find an increase in fontal theta that lateralizes to the right for the competitive condition (CC). While the authors find the lateralization surprising, and instead expected a greater increase over midline structures, the lateralization here may be more consistent with recent literature implicating the right DLPFC in action inhibition. In this case, the inhibited action may be related to the competitor's memory. Alternatively, if there is a conflict signal that is relevant for this task, although the authors look for this conflict signal during retrieval it may also be helpful to identify whether a conflict signal is present during the encoding portion of the task when the competitor memory is introduced.

One of the advantages of the task design is that the subordinate categories allow for classifiers to be built that can decode which memory is being reactivated. The authors use an LDA-based classifier on the MEG sensor amplitudes to construct and test the classifiers. Interestingly, when decoding the retrieval data, the classifiers are significant in the NC conditions a full 2.5 seconds after the cue. This seems like a very long time, as the authors acknowledge. In the description of the task, the authors report that the subjects indicate when they have the picture in mind, and then indicate the supra- and subordinate categories. Presumably, then, this means there is a response time. How does this activation compare to the response time?

The authors report that there is no significant reinstatement of the target category during correct competitive trials as compared to the competitor category, although they claim that the classifier performance peaks at a similar time to the NC condition. This seems a bit concerning because one would expect to see some evidence of reactivation if subjects are making the correct decision. How should this be reconciled? Instead, the authors find that the evidence for target memory increases over recalls, which they offer as evidence that these memories are increasingly being reactivated. However, classifier performance peaks at a different time than when you see reinstatement in NC condition and is in fact earlier. Why?

It would be helpful if the authors could please clarify Supp Figure 3B. What does it mean when they collapse targets and competitors? Do they mean that the classifier can decode up either one?

If the authors are seeing increasing evidence of target information over several repeats, then this raises the question as to whether the classifier of subordinate target category would only work if only looking at the third repeat? The authors test this by comparing classifier performance in the 1st v 3rd repeat. However, how about comparing performance in the 3rd repeat versus chance?

The authors then tie in classifier decoding with the phase of the theta rhythm in the hippocampus. First, the classifier fidelity itself appears to have a 3 Hz rhythm. They then compute the modulation index (MI) of classifier fidelity to the hippocampal 3Hz phase. Classifier fidelity is only modulated by the hippocampal phase during the 3rd retrieval repetition. Similarly, we only see significant phase differences between target and competitor in the last repetition. This would seem to support the hypothesis that repetition leads to greater separation and better memory. However, it would be then helpful to know how the rhythm of classifier fidelity (the underlying 3Hz rhythm of classifier performance) changes across repeats. Is this fixed, or does this also exhibit changes with repetition? Is this also the case in the NC condition? For the NC condition, how do the phases modulation compare with the phases of target and competitors in the CC condition?

https://doi.org/10.7554/eLife.80633.sa1

Author response

Essential revisions:

(1) Concerns were raised about the lack of significant memory reactivations for the CC condition, even at the 3rd repetition when memories are supposed to increasingly be reactivated. The authors collapsed target and competitors (Supp Figure 3B) aiming to support memory reactivations, but the rationale is unconvincing since target and competitors are claimed to occur at different phases and should not be combined for decoding.

We agree that evidence for memory reactivation in the critical CC condition was weak in the previously reported (and preregistered) analyses. We added new analyses showing that robust memory reactivation is present in the CC condition when realigning the timelines to the response (i.e., time point of subjective recollection), rather than the cue onset, which we hope will reassure the readers and reviewers of the validity of our decoding algorithm and approach. We also clarify that the collapsed accuracy for target/competitor decoding only served the particular purpose of identifying a common time window of reactivation.

Regarding the first point of low decoding accuracy in the CC, we reasoned that there is considerable variability in the timing of memory reactivation across trials and participants that will make it difficult to see clear peaks of decoding accuracy when averaging across participants. The revised manuscript now includes several analyses where we realigned decoding to the button press that subjects made on each trial to indicate they recalled the associated memory. This approach (see Author response image 1) revealed significant clusters of classification accuracy preceding the button press by approx. 200ms. In Author response image 1B, we also show that target classification on correct trials is significantly stronger than on incorrect trials, indicating that our method of detecting memory reinstatement is sensitive to behavioural memory success. Lastly, we also find evidence (though not surviving cluster-correction) of competitor reactivation when only using incorrect trials where the competing memory is more likely to dominate, both when doing the analysis response-locked (Author response image 1C) and cue-locked (Figure 2 – supplement 2D; Author response image 1D), the former again in a similar time window around 200ms before response.

Author response image 1
Decoding performance time-locked to the subjective recollection button press.

We realigned the trials based on response a. We find significant target decoding when averaging over all repetitions for targets in the competitive condition (CC), pcluster <.05. b. A similar pattern is evident when contrasting target decoding between correct and incorrect trials in the CC (pcluster <.05). These results reaffirm that the lack of significant decoding in the cue-locked analyses is due to timing differences in memory reinstatement between trials and participants, which is rectified when locking to the time point of subjective recollection. c. When only analysing incorrect trials, we find evidence of competitor decoding in a similar time window as when target memories were maximally reactivated in (a and b) (puncorrected = .05). d. Lastly, analysing competitor decoding on all incorrect trials from cue onset, we again found evidence for competitor reactivation in an early and a later time window (puncorrected = .05).

It is important to emphasise that we preregistered our decoding analyses locked to cue onset in order to maximise our chances to observe theta-locked memory reactivations. Previous work showed that the phase of theta oscillations is reset by a memory cue (Rizzuto et al., 2006; ter Wal et al., 2021), suggesting that the locking of memory reactivations to the theta rhythm will be most coherent for a period of time following cue onset. Since we conducted the crucial analyses of classifier-to-theta locking on the level of single trial fidelity values, comparing phase shifts of target and competitor, the analyses will not be affected by whether or not these reactivation peaks are consistent in time across participants, as they would need to be to reveal robust average cue-locked decoding performance.

Regarding the analysis collapsing of target and competitor decoding, as seen in Supp Figure 3B [now Supp Figure 2E], we understand that this point needs clarification. This analysis was conducted purely for the purpose of obtaining an unbiased time window (i.e., not biased either target or competitor decoding, as suggested as good practice in Cohen (2014)) in which further analyses on the up- and down-regulation of the two memories across repetitions could then be performed (shown in Figure 2 – supplement 2E and Figure 2H). We clarify this rationale in a sentence on p.49 l. 8-12.

For all other analyses, target and competitor decoding was kept separate.

(2) Related to the 1st point, by comparing reactivations of the 3rd and 1st repetition, the authors suggest a reactivation increase over repetitions. The authors could also consider comparing the 3rd repeat and chance level to examine the memory reactivation at the 3rd repetition. Moreover, how do they reconcile the different reactivation latencies for CC and NC conditions?

We followed this suggestion and compared the third repetition of target decoding with chance (50%). The analysis revealed a significant cluster emerging from 1.81 to 2.08 seconds after cue onset. However, this cluster did not survive a more stringent cluster-based permutation correction for multiple comparisons across time (pcorr = .07). We added this information on page p.13 l. 9-10 of the revised manuscript.

Regarding the difference in reactivation latencies, we agree that somewhat surprisingly, decoding performance peaked earlier in the CC than in the NC condition when using the contrast of 3rd vs 1st repetition in the CC condition. Note, however, that decoding of the target memory in the NC and the CC condition (Figure 2D compared to Figure 2E blue) followed a very similar timecourse overall when comparing against chance. Moreover, there is an earlier peak around 1-1.5 seconds in the NC condition that does not reach significance when using a stringent cluster-based permutation analysis, and the cluster-corrected results are thus not entirely conclusive. Also see our response to point #4 of reviewer #3.

(3) Regarding the theta rhythm in classifier fidelity and its relationship to midline theta-band signal, the current results support the phase separation only at the 3rd repetition. It would be important to see how the rhythm of classifier fidelity changes over repeats as well as its characteristics for the NC condition. Moreover, is there any way to confirm that the theta-band rhythm emerges from the hippocampus?

We agree that these are important points to clarify with further analyses. To address the first point about the dominant frequency of memory reactivation in the various conditions, we used IRASA, following the same procedure as in the manuscript, on the fidelity timecourses in the NC and the CC conditions, and separated them into individual recall repetitions. The resulting spectra of the fidelity values are shown in Figure 3 – supplement 1. Reassuringly, the frequency profile and peak frequency did not show an obvious change across repetitions in either condition. However, the NC condition has a slightly faster frequency profile than the CC, with a peak around 4-5Hz (NC) compared with 3 Hz (CC). On p.15 l. 4-11 of the revised manuscript, we briefly discuss that the faster rhythm in the NC compared to the CC condition could reflect an increase in memory load in the CC condition, with participants keeping two memories in mind, consistent with reports of a corresponding slow-down of the theta frequency in the working memory literature (Wolinski et al., 2018).

To more thoroughly analyse the regional specificity of the hippocampal theta and to validate our source analysis, we conducted an additional analysis at source level where we compare the frequency spectrum of our hippocampal region of interest (ROI) with two control areas (superior occipital lobe and primary motor cortex) that should not show memory-related theta oscillations. Figure 3 – supplement 2C-D shows the difference in the raw power spectrum between the hippocampal ROI and each of the two control regions, averaged across 0-2 seconds after cue onset, and showing power differences at frequencies from 1 to 30Hz for visualisation purposes. To statistically evaluate differences in the theta band, we averaged over 3-8 Hz (the frequency band consistently used throughout the manuscript) and used a one-sided t-test to compare the spectrum against that of each control region (Bonferroni-correcting the p-level for the two repeated comparisons). We find that theta power in the hippocampus is significantly increased compared to superior occipital cortex (t(20) = 2.37, pcorrected = .0279), with a maximal power difference at 4 Hz, and a qualitatively similar but non-significant increase when contrasting power over hippocampal and precentral virtual channels (t(20) = 1.98, pcorrected = .0622), again with a maximum difference at 4Hz. The new results are described on p. 16 l. 6-9 of the revised manuscript.

We hope these two results will assure the reviewers/readers that (1) our source localization approach is able to isolate distinct frequency profiles in different ROIs, with the expected theta increase in the hippocampus; and (2) that the frequency profile of the classifier is very similar across repetitions and thus unlikely to contaminate the phase-modulation results. Of course, intracranial EEG data would be required to unequivocally pinpoint the exact source(s) of the 2-4Hz theta rhythm.

(4) There are other behavioral analyses that could be done to relate the neural findings to the behavior in addition to the intrusion effect (for example: probability of intrusion, RT for memory recall, overall accuracy, etc.). Please see details below in the reviewers' comments.

We thank the reviewers for pointing out several interesting analyses that provided further understanding of the results. In response to reviewer 2’s point 3, we conducted an analysis relating neural competitor reactivation to the level of behavioural intrusions. We extracted the average decoding accuracy of competitor memories across repetitions from 0 to 1 second after cue onset (following Newman and Norman, (2010) as suggested by the reviewer) and correlated this neural index with the probability of intrusions in the first repetition. A significant positive correlation (r(1,20) = .446, p = .04) was found indicating that stronger neural reactivation of the competitor memory was related to a higher probability of intrusions. See p.13 l. 23 – p. 14 l. 4. This new finding also strengthens the point that classification performance is meaningfully related to behaviour (see Essential Revisions #1 above).

Second, in response to reviewer 3’s point #1, we further investigated potential changes in behavioural accuracy across repetitions in the CC condition, and found no significant change, (Z = -.73, p = .47; Wilcoxon signed-rank test of linear slope against zero). This result is in line with the previously reported result on the intrusion score and is now reported in the revised manuscript on p.8 l.2-5 and as a Figure 2 – supplement 1A.

Lastly, in response to reviewer 3’s point #3, we analysed the timing of the subjective button presses and how it changes across repetitions. Response times significantly shortened in all conditions (NC2: rep1: 2.01 sec, rep2: 1.75 sec, rep3:1.68 sec; Z = -4.17, p <.01; CC: rep1: 2.29 sec, rep2: 2.05 sec, rep3:2.04 sec; Z = -2.89, p = .004; NC1: rep1: 2.55 sec, rep2: 2.37 sec, rep3:2.26 sec; Z = 3.03, p = .003; z-values are based on Wilcoxon signed-rank tests of linear slope against zero). These results are mainly included for descriptive purposes and show that response times roughly coincided with the temporal latency of our decoding timecourses. We report this response time analysis on p.8 l. 10-16 of the revised manuscript.

(5) The authors could consider adding representative data examples to illustrate the phase separation findings.

In response to this comment, and for transparency, we now show in Figure 4 – supplement 1 each participant’s fidelity timecourse of target decoding relative to competitor decoding, filtered at the relevant 3Hz frequency. Visual inspection shows that target and competitor decoding is often phase shifted and does generally not tend to go hand in hand over the average trial timecourse. To quantify this phase shift, we subtracted, for each participant and time bin, the phase angle of the target from the angle of the competitor (in complex space). The resulting phase separation is highly consistent with the one we reported in the original manuscript when relating the peak fidelity values to the hippocampal 3Hz phase.

We also conducted an additional cross-correlation analysis on the fidelity timecourses of target and competitor memories, to further corroborate the phase separation on a single-trial level. This analysis revealed a significant and temporally extended cluster of phase lag starting approximately 1 sec after cue onset and lasting until the end of the trial, suggesting that competitor reactivation followed target reactivation by a lag of on average 30 degrees on the 3Hz theta cycle. These new findings can be found in revised Figure 4C of the main manuscript.

We hope that together, these illustrative changes and complementary analyses will provide further understanding of the phase separation findings. Note that the fidelity timecourses in the original manuscript (previously Figure 1D) served a purely illustrative purpose and showed simulated not real data. We understand that this illustration was more confusing than helpful, and changed this figure panel to more clearly represent the predictions based on the Norman et al., (2006) model of interference resolution.

Reviewer #1 (Recommendations for the authors):

1. The decoding performance didn't reach statistical significance for both target and competitor (Figure 2e), which makes me skeptical to what extent the decoding courses truly denote the memory reactivations. If not, how could the phase-based memory reactivation conclusion be supported? I noticed that the effect mainly occurred at the 3rd repetition recall. Thus, I would recommend that the authors should at least provide evidence that the decoding performance was significant for the target and competitor in the 3rd repetition.

We agree with the reviewer and now include additional evidence demonstrating that our decoding algorithm is able to pick up memory reactivations with significant above-chance accuracy, including in the competitive condition. Note that while this analysis was not pre-registered, it will hopefully assure the reviewers/readers that we are looking at meaningful reactivation peaks when relating classifier performance to theta phase. In the new analysis, we realigned the decoding timelines to the time point of subjective recollection (i.e., button press to indicate the memory was recalled), instead of cue onset as for all pre-registered analyses. When doing so (see Author response image 1A), we do find significant clusters of target classification accuracy preceding the button press by approx. 200ms. In Author response image 1B, we also show that target classification on correct trials is significantly stronger than on incorrect trials, indicating that our method of detecting memory reinstatement is sensitive to behavioural memory success. Lastly, we also find evidence (though not surviving cluster-correction) of competitor reactivation when only using incorrect trials, both response-locked (Author response image 1C) and cue-locked (Author response image 1D), the former again in a similar time window around 200ms before response.

The most likely reason why a cue-locked decoding approach produces less robust decoding accuracy is the considerable variance in the timing of memory recall (and thus neural pattern reinstatement) across trials, conditions, and participants (see also Reviewer #3 point #3, where we show that RTs differ across repetitions). Such temporal variability is rectified to some degree when aligning the timeline to the button press. Having said that, there are several reasons for why we pre-registered our analyses locked to the onset of the cue. The most important reason is that the phase of theta oscillations is thought to be reset by a retrieval cue (see: Rizzuto et al., 2006; ter Wal et al., 2021), and the locking of memory reactivations to the theta rhythm can thus be expected to be most coherent when aligned to cue onset. Importantly, we conducted the crucial analyses of classifier-to-theta locking on the level of single trial classification peaks, comparing phase shifts of target and competitor peaks. These analyses will thus not be affected by whether or not the reactivation peaks are consistent in time across trials and participants, as they would need to be to reveal robust average cue-locked decoding performance. We added the response-locked target decoding to main Figure 2 (panel g,h) in the revised manuscript, and competitor decoding to Figure 2 – supplement 2D; the description of the new results can be found in the main text on p. 12 l. 15 – 23.

2. The target and competitor showed similar profiles instead of an out-of-phase manner (Figure 1d). Their average decoding performance even showed significant reactivation (supplemental figure 3b). The results seem to be not consistent with the major phase separation findings. Moreover, to flesh out the phase separation results, I would recommend the authors provide typical data examples from their results explicitly illustrating the phase separation profiles, e.g., a representative subject, etc.

We thank the reviewer for pointing out these apparent contradictions. First, it is important to note that Figure 1D served a purely illustrative purpose and shows simulated not real data. In the revised figure, we have now made this panel easier to interpret. Also note that, following the same reasoning as in our response to point #1, a trial-by-trial phase shift between target and competitor reactivation peaks would not necessarily be visible in the cross-subject average of decoding performance, because the timing and absolute phase of target and competitor reactivation can vastly differ in each participant. The average shown in the previous manuscript can only roughly indicate the approximate time window where decoding tends to increase when averaging across all subjects.

To flesh out the phase separation results, we followed the reviewer’s recommendation and now show each individual’s fidelity time course of target and competitor decoding in Figure 4 – supplement. Visual inspection shows that target and competitor reactivation is temporally shifted in most participants, often even showing phase opposition. We also conducted a set of new analyses to quantify this phase shift at the level of single trials and single participants. First, we filtered each participant’s decoding fidelity time course at the dominant 3Hz frequency, separately for target and competitor decoding in the 3rd repetition of the competitive condition. We then subtracted, for each time bin, the phase angle of the target from the angle of the competitor (in complex space), and averaged the phase difference on a group level. The resulting phase separation looks very similar to what we report in the original manuscript when relating the peak fidelity to the hippocampal 3Hz phase. Another way of quantifying the phase lag is to compute a cross-correlation between the two signals, which we did using a sliding-window (to preserve some time information) and calculating the cross-correlation for each 330ms time bin. This analysis revealed a significant cross-correlation cluster starting around 1 sec after cue onset and lasting until the end of the trial, with a maximum lag of 30 degrees (Figure 4C). Together, the two new analyses are highly consistent with the phase shift reported in our pre-registered analysis.

3. The key hypothesis is that the neural representation of the target would become stronger, and phase separation would become more prominent across repetitions. They did find stronger reactivation and phase separation for the 3rd repetition compared to the 1st repetition. Meanwhile, is there any corresponding behavioral evidence such that the overlapping memories become less distracting across repetitions?

This was indeed one of the pre-registered behavioural contrasts and reported on p.8 l.1 of the original manuscript. Contrary to our predictions, we found no significant difference in intrusion scores across repetitions, and this relatively stable pattern of intrusions across repeats might be due to the lack of feedback in our task, as discussed on p.23 l.19-22 (also see Reviewer 3’s comment 2; Figure 5).

4. There is no test to confirm the significant theta rhythm in the hippocampus detected in the present study. I think the authors should provide neural evidence backing up that the theta-band rhythm analyzed in their study indeed derives from the hippocampus.

We agree that the original manuscript was lacking evidence for the specificity of the hippocampal theta. We conducted a new analysis to address this concern, computing the frequency spectrum of the raw (trial) signals extracted from virtual sensors in the hippocampus, and comparing it to the spectrum of two control areas (superior occipital lobe and primary motor cortex) that should not show memory-related theta oscillations. Figure 3 – supplement 2 shows the difference in the power spectrum between the hippocampal ROI and each of the control regions, averaged across 0-2 seconds after cue onset, and showing power differences at frequencies from 1 to 30Hz for visualisation purposes. To statistically evaluate differences in the theta band, we averaged power over 3-8 Hz (the band consistently used throughout the manuscript). We find that theta power in the hippocampus is significantly increased compared to superior occipital cortex (t(20) = 2.37, pcorrected = .0279), with a maximal power difference at 4 Hz. A qualitatively similar but non-significant increase was found when contrasting the spectrum from hippocampal and precentral virtual channels (t(20) = 1.98, pcorrected = .0622), again with a maximum difference at 4Hz. We hope these results will assure the reviewers/readers that our source localization approach is able to isolate distinct frequency profiles in different ROIs, with the expected theta increase in the hippocampus. We added the new results to the result section (p.16 l.5-8) and added two panels to Figure 3 – supplement 2. Of course, MEG source localisation is still an imperfect reconstruction, and we emphasize that while we can show some degree of specificity, intracranial EEG data would be required to unequivocally locate the main source(s) of the 2-4Hz theta rhythm.

Reviewer #2 (Recommendations for the authors):

I found much to appreciate about this manuscript and work. Much of that is summarized in my 'Public review' so I won't repeat it here.

I had no 'substantial concerns' about the quality of the work or interpretations presented. That is to say, none of the following thoughts that I'll share next should be seen as show-stopping. I will share these thoughts nonetheless in case they can support the growth of this work.

– The reasons why the reactivation effects should only become visible in the 3rd repetition were not clear to me. I struggle to come up with a just-so story based on the Norman et al. 2006 model or otherwise to understand these. Some assistance with this could help.

We appreciate this comment and clarify our reasoning in the revised manuscript (p. 4 l.8-21). First, it should be highlighted that in our pre-registration, we hypothesized (i) an increase in target and a decrease in competitor reactivation (i.e., decodability) over repetitions; and (ii) a target-competitor phase difference that exists throughout the task and would increase across repetitions. These hypotheses were only partly confirmed, and we hope we are sufficiently transparent about this in the manuscript. The first hypothesis was motivated by a wealth of empirical studies on retrieval-induced enhancement of target memories and retrieval-induced forgetting of competitor memories (Anderson et al., 1994; Wimber et al., 2015). It can also be deduced indirectly from the Norman et al., (2006) model, which assumes that over time and repetitions of the presumed theta cycle dynamics, weak target nodes are strengthened while overly strong competitor nodes are weakened, leading to better differentiation hence enhanced decodability of the target memory. We now visualise this reasoning in the revised hypothesis graph in Figure 1D, where targets become decodable at increasingly early (higher inhibition) theta phases with gradually less interference from competitors.

The second hypothesis of increasing phase separation over time is based directly on the model’s dynamics. Weak target nodes do not survive high inhibition initially and thus only activate during medium and lower inhibition phases of the theta cycle, however, with repeated strengthening they will become active at an increasingly early, higher-inhibition phase. Therefore, while early in time the target and competitor memories overlap in their peak reactivation phase, this overlap gets reduced by the strengthening and weakening dynamics in the model. Critically, Norman et al. (2006) model these dynamics across several repetitions of a theta cycle, while we instead opted to measure them across recall repetitions, simply because we were not confident that our analysis tools could provide a stable phase estimate for target and competitor reactivation on each theta cycle within a retrieval trial.

– There is a strange discrepancy in the frequency band of the frontal midline theta (~7 Hz) and the frequency of the phase modulated reactivations. I did not see any attempt to reconcile this. Are they simply two totally different things, the 7Hz, and 3Hz thetas?

This is a valid observation and indeed something we did not discuss in the original manuscript. To our knowledge, the relationship between the conflict-related frontal midline theta and the memory-related 2-4Hz theta is still unknown, though both rhythms are consistently found in their respective cognitive domains. The frontal midline theta is typically found at 7-8Hz, like in our own study, and has been related to cognitive conflict/control across domains (Cavanagh and Frank, 2014). In the memory domain, frontal midline theta amplitudes increase with increased competition and decrease with downregulation of the competitor memory (Ferreira et al., 2014; Hanslmayr et al., 2010). Our own finding that power in this frontal theta rhythm increases under conditions of high competition is thus a replication of previous work, and a sanity check that physiological markers of competition increase in the CC compared to the NC condition. A memory-related theta in the human hippocampus is frequently found around 2-3Hz in intracranial recordings (Goyal et al., 2020; Lega et al., 2012; ter Wal et al., 2021), and thus slower than the dominant theta rhythm in the rodent hippocampus. Our findings suggest that the phase separation of competing memories relates to this slower hippocampal rhythm, not the faster frontal rhythm. Beyond this observation, we are not aware of literature linking the two theta rhythms directly, though this would certainly be an interesting field for future studies. Hence, for the time being, we treat these two rhythms as separate phenomena, and clarify our reasoning on p.9 l.11-14 of the revised manuscript.

– There are other tests that could be done to relate the reactivation dynamics to the behavior than are described. For example, one could compare the degree of reactivation to the probability of intrusion. Newman et al., 2010 – https://pubmed.ncbi.nlm.nih.gov/20181622/ – for example did something similar. At the coarsest level, the probability of an intrusion should be lowest for trials where there was no evidence of competitor reactivation.

Thank you for this suggestion. Taking inspiration from the suggested method, we extracted the average decoding across repetitions for competitor memories (using the same trials as in the manuscript) from 0 to 1 second after cue onset (following Newman and Norman, (2010)) and correlated this with the probability of intrusions in the first repetition. This yielded a significant correlation (r(1,20) = .446, p = .04), such that stronger reactivation of the competitor memory was associated with a higher probability of intrusions. This new finding also adds evidence in support of the decoding performance being related to behaviour. See Figure 2 – supplement 1C and p.13 l. 23 – p.14 l.4, of the revised manuscript.

Reviewer #3 (Recommendations for the authors):

There are some suggestions that could improve the overall conclusions one can draw from the manuscript.

The manipulation of repeating the retrieval period is interesting and allows for some novel hypotheses and questions. The authors use a measure of intrusions (by assessing how often subjects select the competitor subordinate category) and find that this is independent of the recall trial number. How about accuracy for the CC conditions?

We thank the reviewer for this suggestion. We did not find a change in behavioural accuracy of CC across repetitions, (Z = -.73, p = .47; Wilcoxon signed-rank test of linear slope against zero). This additional data point is now reported in the revised manuscript on p.8 l.2-5 and as a Figure 2 – supplement 1A. Note that while intrusion scores remain at the same level over recall repetitions, we report new analyses showing that the intrusions do relate to behaviour in a sensible way, such that more neural competitor activation is associated with a higher likelihood of intrusions, see response to Reviewer #2’s point 3.

The authors find an increase in fontal theta that lateralizes to the right for the competitive condition (CC). While the authors find the lateralization surprising, and instead expected a greater increase over midline structures, the lateralization here may be more consistent with recent literature implicating the right DLPFC in action inhibition. In this case, the inhibited action may be related to the competitor's memory. Alternatively, if there is a conflict signal that is relevant for this task, although the authors look for this conflict signal during retrieval it may also be helpful to identify whether a conflict signal is present during the encoding portion of the task when the competitor memory is introduced.

We thank the reviewer for these comments and address them as follows. Regarding the right lateralization of the ‘conflict theta’, we added a section to the main results (p. 9 l. 9-11) to highlight that response conflict and response inhibition are indeed often found to engage a right-lateralised frontal network. Regarding the conflict signal during the encoding phase, we agree that a thorough analysis of the memory (re)activation processes occurring already during encoding is very interesting and we are planning such analyses in the future. We do think that a full analysis of the encoding data is beyond the scope of the current paper. Having said that, in response to this reviewer comment, we analysed theta power when participants encode a second competing memory (CC condition) in the learning phase, compared to encoding the same associate a second time. This analysis indeed reveals a non-significant theta power increase in the CC condition with a maximum around 100-300ms after second picture onset, which could be interpreted as conflict signal (see Author response image 2). Note, however, that during encoding, this contrast is naturally confounded with repetition (NC2) vs no-repetition (CC) of an image, and can therefore not isolate the pure competition component. We decided to not include these analyses in the revised manuscript for the time being, although we would be happy to include them in the supplements if the reviewers/editors consider them central to the conclusions.

Author response image 2
Conflict theta during second presentation at encoding.

We found that there is stronger theta power when participants encode a second competing associate (target item) with an old memory cue in the CC condition, compared to re-encoding the first associate with the old memory cue in the NC condition. X-axis shows time from image onset (0) and one second after the image appeared on screen.

One of the advantages of the task design is that the subordinate categories allow for classifiers to be built that can decode which memory is being reactivated. The authors use an LDA-based classifier on the MEG sensor amplitudes to construct and test the classifiers. Interestingly, when decoding the retrieval data, the classifiers are significant in the NC conditions a full 2.5 seconds after the cue. This seems like a very long time, as the authors acknowledge. In the description of the task, the authors report that the subjects indicate when they have the picture in mind, and then indicate the supra- and subordinate categories. Presumably, then, this means there is a response time. How does this activation compare to the response time?

We thank the reviewer for highlighting this observation, and we further investigated the temporal relationship between memory reactivation and response times in a number of new analyses. Response times for the subjective recollection button press were on average 2.11 sec, and the peak decodability therefore indeed coincides roughly with the response. Response times significantly shortened in all conditions across the three repetitions (NC2: rep1: 2.01 sec, rep2: 1.75 sec, rep3:1.68 sec; Z = -4.17, p <.01; CC: rep1: 2.29 sec, rep2: 2.05 sec, rep3:2.04 sec; Z = -2.89, p = .004; NC1: rep1: 2.55 sec, rep2: 2.37 sec, rep3:2.26 sec; Z = 3.03, p = .003; Wilcoxon signed-rank tests of linear slope against zero). In response to this reviewer comment and Reviewer #1’s point #1 (see Author response image 2 and the above response), we performed an additional analysis locking the decoding to response onset rather than cue onset. This analysis reveals significant clusters of memory reactivation (i.e. decoding peaks) approximately 200ms before the subjective recollection response. We report these new analyses on p.8 l. 10-16 of the revised manuscript.

The authors report that there is no significant reinstatement of the target category during correct competitive trials as compared to the competitor category, although they claim that the classifier performance peaks at a similar time to the NC condition. This seems a bit concerning because one would expect to see some evidence of reactivation if subjects are making the correct decision. How should this be reconciled? Instead, the authors find that the evidence for target memory increases over recalls, which they offer as evidence that these memories are increasingly being reactivated. However, classifier performance peaks at a different time than when you see reinstatement in NC condition and is in fact earlier. Why?

We appreciate that the evidence for target reactivation in the CC was not sufficiently strong based on the (pre-registered) analyses reported previously. As the reviewer also points out, decoding performance peaked earlier in the CC than in the NC condition, at least when based on the comparison of 3rd vs 1st repetition (where on the 3rd repetition, memory recall might indeed have happened earlier than on the 1st repetition). When comparing target decoding against chance level in the NC and CC conditions (Figure 2D and 2E), they do follow very similar timecourses. Moreover, an earlier peak around 1-1.5 seconds can also be seen in the NC condition but does not reach significance when using a stringent cluster-based permutation statistic, and the results are thus not entirely conclusive. Addressing the concern about general decoding performance and relationship to behaviour, we conducted the above-mentioned response-locked analyses, showing a clear reactivation peak just before the response when aligning trials to the time point of subjective recollection. Moreover, and relevant to the specific point raised here, we also show that this target reactivation is significantly stronger when participants make a correct vs incorrect decision, suggesting there is more neural evidence for the target class when participants correctly retrieve this class. More generally, response-locked reactivation appears to be more robust suggesting to us that there is considerable variance in the timing of memory recall across trials, conditions, and participants (also see Essential Revisions point #1, Reviewer #1 point #1, and Reviewer #3 point #3), and such variability makes it difficult to see clear cue-locked reactivation peaks on. Corresponding changes in the manuscript occur in the revised Figure 2 and in the main text on p. 12 l. (15 – 23).

It would be helpful if the authors could please clarify Supp Figure 3B. What does it mean when they collapse targets and competitors? Do they mean that the classifier can decode up either one?

We understand that this point needs clarification. Of course, the design of our study and all our main analyses are tailored to separate target and competitor reactivation. This specific analysis collapsing across target and competitor reactivation was conducted purely for the purpose of identifying a time window where on average, across all trials and participants, target and competitor reactivation tend to be high, such that we could then use this unbiased time window (i.e., not biased towards target or competitor decoding) to probe for the hypothesized up and down-regulation of target and competitor memories, respectively, across repetitions (as seen in Figure 2 – supplement 2B and Figure 2F). For all other analyses, we did keep target and competitor decoding separate. We clarify this reasoning now in a sentence on p.49 l. 8-12. See also Essential Revisions point #1.

If the authors are seeing increasing evidence of target information over several repeats, then this raises the question as to whether the classifier of subordinate target category would only work if only looking at the third repeat? The authors test this by comparing classifier performance in the 1st v 3rd repeat. However, how about comparing performance in the 3rd repeat versus chance?

We followed the reviewer’s suggestion, and we do also find significant decoding performance for target memories when comparing against chance (50%) in the third repetition, with a significant cluster emerging from 1.81 to 2.08 seconds after cue onset. However, this cluster does not survive a more stringent cluster-based permutation correction for multiple comparisons across time (pcorr = .07). We added this analysis on page p.13 l. 9-10. See also Essential Revisions point #2.

The authors then tie in classifier decoding with the phase of the theta rhythm in the hippocampus. First, the classifier fidelity itself appears to have a 3 Hz rhythm. They then compute the modulation index (MI) of classifier fidelity to the hippocampal 3Hz phase. Classifier fidelity is only modulated by the hippocampal phase during the 3rd retrieval repetition. Similarly, we only see significant phase differences between target and competitor in the last repetition. This would seem to support the hypothesis that repetition leads to greater separation and better memory. However, it would be then helpful to know how the rhythm of classifier fidelity (the underlying 3Hz rhythm of classifier performance) changes across repeats. Is this fixed, or does this also exhibit changes with repetition? Is this also the case in the NC condition? For the NC condition, how do the phases modulation compare with the phases of target and competitors in the CC condition?

We agree that changes in classifier-to-phase locking over time might be ‘contaminated’ by frequency changes. We thus followed the reviewer’s suggestion to analyse the dominant frequency in the fidelity timecourses in both NC and CC across repetitions. Both conditions show peaks in the theta range, though the NC condition has a slightly faster frequency profile than the CC: 4-5Hz (NC) compared with 3 Hz (CC). Reassuringly, there is no obvious change in the peak frequency across repetitions in either condition. We now show these spectra of the fidelity values in Figure 3 – supplement 1. Note that there is a potentially interesting explanation for the faster rhythm in the NC than the CC condition, such that the increased memory load in the CC (with participants keeping two memories in mind) being adapted to by a slow-down of the theta frequency as previously reported in the working memory literature (Wolinski et al., 2018). We briefly allude to this possibility on p.15 l. 4-11 of the revised result section and p.28 l.23 – p.29 l.4 of the discussion. See also Essential Revisions #3.

References

Anderson, M. C., Bjork, R. A., and Bjork, E. L. (1994). Remembering can cause forgetting: Retrieval dynamics in long-term memory. Journal of Experimental Psychology. Learning, Memory, and Cognition, 20(5), 1063–1087. https://doi.org/10.1037//0278-7393.20.5.1063

Cavanagh, J. F., and Frank, M. J. (2014). Frontal theta as a mechanism for cognitive control. Trends in Cognitive Sciences, 18(8), 414–421. https://doi.org/10.1016/j.tics.2014.04.012

Cohen, M. X. (2014). Analyzing Neural Time Series Data: Theory and Practice. MIT Press.

Ferreira, C. S., Marful, A., Staudigl, T., Bajo, T., and Hanslmayr, S. (2014). Medial prefrontal theta oscillations track the time course of interference during selective memory retrieval. Journal of Cognitive Neuroscience, 26(4), 777–791. https://doi.org/10.1162/jocn_a_00523

Goyal, A., Miller, J., Qasim, S. E., Watrous, A. J., Zhang, H., Stein, J. M., Inman, C. S., Gross, R. E., Willie, J. T., Lega, B., Lin, J.-J., Sharan, A., Wu, C., Sperling, M. R., Sheth, S. A., McKhann, G. M., Smith, E. H., Schevon, C., and Jacobs, J. (2020). Functionally distinct high and low theta oscillations in the human hippocampus. Nature Communications, 11(1), 2469. https://doi.org/10.1038/s41467-020-15670-6

Hanslmayr, S., Staudigl, T., Aslan, A., and Bäuml, K.-H. (2010). Theta oscillations predict the detrimental effects of memory retrieval. Cognitive, Affective and Behavioral Neuroscience, 10(3), 329–338. https://doi.org/10.3758/CABN.10.3.329

Kerrén, C., Linde-Domingo, J., Hanslmayr, S., and Wimber, M. (2018). An Optimal Oscillatory Phase for Pattern Reactivation during Memory Retrieval. Current Biology: CB, 28(21), 3383-3392.e6. https://doi.org/10.1016/j.cub.2018.08.065

Lega, B. C., Jacobs, J., and Kahana, M. (2012). Human hippocampal theta oscillations and the formation of episodic memories. Hippocampus, 22(4), 748–761. https://doi.org/10.1002/hipo.20937

Linde-Domingo, J., Treder, M. S., Kerrén, C., and Wimber, M. (2019). Evidence that neural information flow is reversed between object perception and object reconstruction from memory. Nature Communications, 10(1), 179. https://doi.org/10.1038/s41467-018-08080-2

Newman, E. L., and Norman, K. A. (2010). Moderate Excitation Leads to Weakening of Perceptual Representations. Cerebral Cortex (New York, NY), 20(11), 2760–2770. https://doi.org/10.1093/cercor/bhq021

Norman, K. A., Newman, E., Detre, G., and Polyn, S. (2006). How Inhibitory Oscillations Can Train Neural Networks and Punish Competitors. Neural Computation, 18(7), 1577–1610. https://doi.org/10.1162/neco.2006.18.7.1577

Rizzuto, D. S., Madsen, J. R., Bromfield, E. B., Schulze-Bonhage, A., and Kahana, M. J. (2006). Human neocortical oscillations exhibit theta phase differences between encoding and retrieval. NeuroImage, 31(3), 1352–1358. https://doi.org/10.1016/j.neuroimage.2006.01.009

Staresina, B. P., and Wimber, M. (2019). A Neural Chronometry of Memory Recall. Trends in Cognitive Sciences, 23(12), 1071–1085. https://doi.org/10.1016/j.tics.2019.09.011

ter Wal, M., Linde-Domingo, J., Lifanov, J., Roux, F., Kolibius, L. D., Gollwitzer, S., Lang, J., Hamer, H., Rollings, D., Sawlani, V., Chelvarajah, R., Staresina, B., Hanslmayr, S., and Wimber, M. (2021). Theta rhythmicity governs human behavior and hippocampal signals during memory-dependent tasks. Nature Communications, 12(1), 7048. https://doi.org/10.1038/s41467-021-27323-3

Wimber, M., Alink, A., Charest, I., Kriegeskorte, N., and Anderson, M. C. (2015). Retrieval induces adaptive forgetting of competing memories via cortical pattern suppression. Nature Neuroscience, 18(4), 582–589. https://doi.org/10.1038/nn.3973

Wolinski, N., Cooper, N. R., Sauseng, P., and Romei, V. (2018). The speed of parietal theta frequency drives visuospatial working memory capacity. PLOS Biology, 16(3), e2005348. https://doi.org/10.1371/journal.pbio.2005348

https://doi.org/10.7554/eLife.80633.sa2

Article and author information

Author details

  1. Casper Kerrén

    1. Centre for Human Brain Health, School of Psychology, University of Birmingham, Birmingham, United Kingdom
    2. Research Group Adaptive Memory and Decision Making, Max Planck Institute for Human Development, Berlin, Germany
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    casper.kerren@gmail.com
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4870-6072
  2. Sander van Bree

    Centre for Cognitive Neuroimaging, School of Neuroscience and Psychology, University of Glasgow, Glasgow, United Kingdom
    Contribution
    Formal analysis, Methodology, Writing - original draft
    Competing interests
    No competing interests declared
  3. Benjamin J Griffiths

    Centre for Human Brain Health, School of Psychology, University of Birmingham, Birmingham, United Kingdom
    Contribution
    Formal analysis, Methodology, Writing - original draft
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8600-4480
  4. Maria Wimber

    1. Centre for Human Brain Health, School of Psychology, University of Birmingham, Birmingham, United Kingdom
    2. Centre for Cognitive Neuroimaging, School of Neuroscience and Psychology, University of Glasgow, Glasgow, United Kingdom
    Contribution
    Conceptualization, Data curation, Supervision, Funding acquisition, Validation, Methodology, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    Maria.Wimber@glasgow.ac.uk
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1917-353X

Funding

European Research Council (2016-StG-715714)

  • Maria Wimber

Stiftelsen Olle Engkvist Byggmästare (Awarded for Ph.D studies)

  • Casper Kerrén
  • Maria Wimber

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This work was supported by a fellowship from Stiftelsen Olle Engkvist Byggmästare awarded to MW and CK and a Starting Grant from the European Research Council awarded to MW (ERC-2016-StG-715714). We also thank Sebastian Michelmann for helpful conceptual input during data analysis.

Ethics

Human subjects: All experimental procedures in the present study were approved by and conducted in accordance with the University of Birmingham's STEM Research Ethics Committee (ERN-16-1512 and ERN-18-0226P). Written informed consent was obtained from participants before they took part in the experiment.

Senior Editor

  1. Laura L Colgin, University of Texas at Austin, United States

Reviewing Editor

  1. Huan Luo, Peking University, China

Reviewer

  1. Ehren L Newman, Indiana University, United States

Publication history

  1. Preprint posted: May 8, 2022 (view preprint)
  2. Received: May 27, 2022
  3. Accepted: November 3, 2022
  4. Version of Record published: November 17, 2022 (version 1)

Copyright

© 2022, Kerrén et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 393
    Page views
  • 83
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Casper Kerrén
  2. Sander van Bree
  3. Benjamin J Griffiths
  4. Maria Wimber
(2022)
Phase separation of competing memories along the human hippocampal theta rhythm
eLife 11:e80633.
https://doi.org/10.7554/eLife.80633

Further reading

    1. Neuroscience
    Kevin Vinberg, Jörgen Rosén ... Fredrik Ahs
    Research Article Updated

    Understanding the neural basis for individual differences in the skin conductance response (SCR) during discriminative fear conditioning may inform on our understanding of autonomic regulation in fear-related psychopathology. Previous region-of-interest (ROI) analyses have implicated the amygdala in regulating conditioned SCR, but whole brain analyses are lacking. This study examined correlations between individual differences in SCR during discriminative fear conditioning to social stimuli and neural activity throughout the brain, by using data from a large functional magnetic resonance imaging study of twins (N = 285 individuals). Results show that conditioned SCR correlates with activity in the dorsal anterior cingulate cortex/anterior midcingulate cortex, anterior insula, bilateral temporoparietal junction, right frontal operculum, bilateral dorsal premotor cortex, right superior parietal lobe, and midbrain. A ROI analysis additionally showed a positive correlation between amygdala activity and conditioned SCR in line with previous reports. We suggest that the observed whole brain correlates of SCR belong to a large-scale midcingulo-insular network related to salience detection and autonomic-interoceptive processing. Altered activity within this network may underlie individual differences in conditioned SCR and autonomic aspects of psychopathology.

    1. Neuroscience
    Rong Zhao, Stacy D Grunke ... Joanna L Jankowsky
    Research Article

    Neurodegenerative diseases are characterized by selective vulnerability of distinct cell populations; however, the cause for this specificity remains elusive. Here we show that entorhinal cortex layer 2 (EC2) neurons are unusually vulnerable to prolonged neuronal inactivity compared with neighboring regions of the temporal lobe, and that reelin+ stellate cells connecting EC with the hippocampus are preferentially susceptible within the EC2 population. We demonstrate that neuronal death after silencing can be elicited through multiple independent means of activity inhibition, and that preventing synaptic release, either alone or in combination with electrical shunting, is sufficient to elicit silencing-induced degeneration. Finally, we discovered that degeneration following synaptic silencing is governed by competition between active and inactive cells, which is a circuit refinement process traditionally thought to end early in postnatal life. Our data suggests that the developmental window for wholesale circuit plasticity may extend into adulthood for specific brain regions. We speculate that this sustained potential for remodeling by entorhinal neurons may support lifelong memory but renders them vulnerable to prolonged activity changes in disease.