Introduction

Early research suggested that the hippocampus is essential exclusively for long-term episodic memory formation^1–3. However, emerging evidence implicates that the hippocampus also contributes to visual short-term memory (VSTM)^4–7. For example, patients with focal hippocampal damage show impaired VSTM^4,8,9. Moreover, item-specific VSTM representations could be decoded and reconstructed from the hippocampus^10–12. These findings suggest that the hippocampus contributes to maintaining VSTM representations, yet the underlying mechanisms remain unclear.

The current study specifically focuses on the hippocampal ripple—a brief high-frequency oscillation (∼80-150 Hz) that reflects synchronized activity of neuronal ensembles in the hippocampus¹³. Hippocampal ripple rates are initially well established in post-learning sleep, where learning increases ripple occurrence during subsequent slow-wave sleep¹⁴. In addition, enhanced hippocampal ripple rates have also been observed during successful memory encoding and long-term memory retrieval^15–19. In contrast, their role in VSTM is far from clear. Recent rodent studies reported that selectively disrupting hippocampal ripples during awake learning impaired the working memory-dependent spatial navigation performance^20,21, suggesting a possible—but still largely unexplored—contribution of ripples to VSTM.

The traditional persistent neural activity model of short-term memory suggests that memory is actively maintained through the sustained activation of neurons that selectively code memory content²². This view was originally inspired by early non-human primate studies²³ and further supported by some human iEEG studies showing persistently increased neural activity during delay periods in short-term memory tasks^10,11. However, such persistent activity is not consistently observed in the neocortex, where the short-term memory information can be successfully decoded^24,25. Moreover, hippocampal activity during maintenance is often at or below baseline for low memory loads, with increases under higher loads²⁶. These findings motivated alternative dynamic coding models, which posit that VSTM representations are maintained not through sustained firing, but via short-term synaptic plasticity that stores information in an “activity-silent” hidden state^27–30. According to these models, transient bursts are required to refresh the short-lived latent synaptic trace and thus maintain the stored representations. Moreover, these bursts are posited to “ramp up” toward the end of the retention interval, reflecting memory readout and preparation for responses³⁰, a prediction supported by non-human primate research^31,32.

Hippocampal ripples share key properties with these postulated reactivation bursts. They coordinate temporally compressed replay of neural firing patterns within the hippocampus¹³ and are implicated in facilitating the transfer of these compressed hippocampal representations to distributed cortical networks during sleep-based memory consolidation^33–36. Such hippocampal-neocortical coordination can be achieved via coupled ripple, i.e., co-occurrence of hippocampal ripples and neocortical ripples³⁷, which have also been associated with successful memory retrieval^18,38. In particular, we are interested in the interaction between the hippocampus and the lateral temporal lobe (LTL), a region implicated in representing memory-specific content during episodic memory retrieval^18,39. These findings raise a key question: Do hippocampal ripples act as transient reactivation bursts that coordinate LTL activity to support human VSTM?

To answer this question, we analyzed iEEG recordings from the hippocampus (HPC) and lateral temporal lobe (LTL) in neurosurgical patients performing a VSTM task that required maintaining and then discriminating target images from highly similar lures. We found that hippocampal but not LTL ripple rates ramped up during the maintenance period, supporting VSTM accuracy. Critically, hippocampal ripples were temporally coupled with LTL ripples, and these coupled ripples further coincided with memory reactivation in the LTL. Importantly, these VSTM-related ripple dynamics cannot be attributed to formation of subsequent long-term memory. Together, these findings provide direct evidence that hippocampal ripples coordinate neocortical memory reactivation through dynamic coding mechanisms in supporting the VSTM.

Results

Behavioral results

Thirteen participants (mean age ±SD: 26.77 ±5.48 years, 7 females) with drug-resistant epilepsy were included in this study. The VSTM was assessed using a delayed match-to-sample (DMS) task comprising three stages: encoding, maintenance, and retrieval (Fig. 1a, upper; see STAR Methods). In each trial, participants encoded a word-picture pair for 3 s. This was followed by a 7-second maintenance period, during which the picture was removed and participants were asked to vividly mentally maintain the picture. During the immediate retrieval stage, a probe picture—either the original target or a highly similar lure—was presented, and participants indicated whether it matched the encoded picture (see Fig. 1a lower). All pictures were from four categories (i.e., animals, fruits, electrical devices, and furniture). Participants performed well in the DMS task with an accuracy of 89.71 ±4.26% (mean ±SD) and a response time (RT) of 1.03 ±0.22 s (mean ±SD). In addition, the RT for correct trials (i.e., remembered trials) was significantly shorter than that for incorrect trials (i.e., forgotten trials, t(12) = −5.88, p < 0.001). Note that the cue word in the DMS task was designed for a subsequent long-term memory cued-recall test, which enabled us to separate VSTM-related neural dynamics from those associated with long-term memory formation.

Figure 1.

Hippocampal ripple rates during the VSTM task

All participants were implanted with depth electrodes for clinical purposes. The iEEG data were recorded from the hippocampus during the DMS task (69 channels, mean ±SD: 5.31 ± 5.01 channels per patient; Fig. 1b). To investigate whether hippocampal ripples are modulated by VSTM, we first extracted ripples from individual hippocampal channels following the well-established protocols in the previous studies^16,17,40. Specifically, raw iEEG data were bipolar re-referenced and filtered between 70 and 180 Hz (see STAR Methods). Then, the amplitude of the data was computed using a Hilbert transform, which was further rectified, squared, smoothed, and converted to z-scores. Ripple events were defined as transient amplitude fluctuations exceeding 4 SD above the pre-encoding baseline (i.e., 200-800 ms before stimulus onset), with durations in the range of 20 - 200 ms (Fig. 2a, see STAR Methods).

Figure 2.

To validate our ripple detection procedures, we computed the ripple rate (i.e., number of ripple events per second) for individual hippocampal channels (see Fig. 2b upper panel for discrete ripples for individual trials from one exemplar channel). Across all task stages and participants, the mean ripple rate was 0.29 event/sec (Hz) (see Fig. S1), in line with previous studies^16,17,40. Moreover, we replicated the novelty effect on ripple rates during memory encoding¹⁶. Specifically, hippocampal ripple rates were significantly higher for novel trials (i.e., the first presentation of VSTM items) than repeated trials (i.e., the second and third presentations of VSTM items) during encoding (p_cluster < 0.001, corrected by the non-parametric cluster-based permutation tests, see Fig. S2).

We next examined ripple rate change across different stages—encoding, maintenance, and retrieval. Hippocampal ripples occurred in 54.26% (SD = 13.55%) of trials during encoding, 78.74% (SD = 13.25%) during maintenance, and 26.85% (SD = 11.27%) during the pre-retrieval stage. Ripple rates were first averaged within each stage per trial. To account for trial-level variability, we then analyzed ripple rate changes across stages using a linear mixed-effects model with task stage as a fixed effect, with participant and trial as random effects (see STAR Methods). Our results showed that, compared to the pre-encoding baseline (i.e., −800 to −200 ms relative to stimulus onset), hippocampal ripple rates significantly increased during encoding and retrieval (all ps_FDR < 0.009, multiple comparisons corrected by false discovery rate (FDR)) but not during maintenance (β = 0.141, z = 2.462, p_FDR = 0.079, Fig. 2c). Direct comparisons showed higher ripple rates during retrieval than during encoding (β = 0.042, z = 7.203, p_FDR < 0.001) and maintenance (β = 0.046, z = 8.018, p_FDR < 0.001), whereas encoding and maintenance did not differ significantly (β = 0.005, z = 0.822, p_FDR = 0.844).

To further examine ripple dynamics at a finer time scale, we performed a time-resolved analysis by calculating the hippocampal ripple rates within consecutive 100 ms non-overlap sliding time windows for each task stage, averaging across channels for each participant. These ripple rates were then compared to the pre-encoding baseline across participants using cluster-based permutation tests⁴¹. The analysis revealed significant clusters showing increased hippocampal ripple rates during early encoding (0.15-0.85 s post stimulus-onset, p_cluster = 0.018; Fig. 2b) and the late maintenance (7.85-8.15 s post stimulus-onset and 8.65-9.45 s post-stimulus onset, ps_cluster < 0.034). In contrast, ripple rates decreased significantly immediately after the response (0.05 s prior to 1.0 s post retrieval response, p_cluster = 0.001), consistent with the decrease observed following long-term memory retrieval in previous studies^17,40.

We next performed the same analyses on iEEG data from the lateral temporal lobe (LTL) in the same participants (132 channels; mean ±SD: 10.15 ±7.82 channels per patient; Fig. 1b). On average, LTL ripples occurred in 32.17% (SD = 14.69%) of trials during encoding, 52.34% (SD = 17.82%) during maintenance, and 14.99% (SD = 10.30%) during the pre-retrieval stage. Similar to the hippocampus, LTL ripple rates during encoding and retrieval were significantly higher than the pre-encoding baseline (both ps_FDR < 0.002), whereas maintenance did not differ from baseline (β = −0.008, z = −2.505, p_FDR = 0.059, Fig. 2d-e). Ripple rates during both encoding and retrieval were greater than during maintenance (both ps_FDR < 0.001), and retrieval showed higher ripple rates than encoding (β = 0.033, z = 10.436, p_FDR < 0.001). The time-resolved group-level analysis revealed a significant cluster of increased ripple activity during early encoding (0.15-0.55 s post stimulus onset, p_cluster = 0.047; Fig. 2d).

To assess whether hippocampal ripple rates were associated with VSTM performance, we first compared ripple rates between remembered and forgotten trials (i.e., memory accuracy), as well as between fast and slow responses among remembered trials (i.e., retrieval speed). Replicating previous studies¹⁷, hippocampal ripple rates are higher for successful than unsuccessful long-term memory retrieval (Fig. S2). However, no significant differences were observed between remembered and forgotten trials for any VSTM stages (all ps_cluster > 0.466; Fig. S2). Besides, fast and slow trials were defined as lower or higher than each participant’s median reaction time (RT), respectively. Faster trials exhibited significantly higher ripple rates than slow trials during the pre-response retrieval (p_cluster = 0.014; Fig. S2). Same analyses on LTL channels revealed no significant effects for either memory accuracy or retrieval speed (ps_cluster > 0.237, Fig. S2).

These findings together suggest that while both hippocampal and LTL ripple rates were modulated by the VSTM task, these ripple rates were not associated with VSTM performance. Besides, for the hippocampal channels, ripple rates were above pre-encoding baseline during the late maintenance time windows, suggesting a potential ramping-up pattern that was tested in the following analyses.

Ramping-up hippocampal ripples during maintenance supports successful VSTM

Next, we tested whether hippocampal ripples ramp up during the maintenance stage, as proposed by the dynamic coding framework³⁰. Using a logistic generalized linear mixed-effects model (GLMM) with ripple occurrence as the dependent variable, maintenance time as fixed effects, and random intercepts/slopes for participants and trials, we observed a significant fixed effect of time across all trials (β = 0.012, z = 4.678, p < 0.001, Fig. S3), suggesting a ramping-up effect of hippocampal ripples. To examine whether this ramping-up predicted the VSTM performance, we further fitted a similar GLMM model with the interaction term: maintenance time × VSTM accuracy (i.e., remember versus forget) as fixed effects (see STAR Methods). Our results revealed a significant interaction effect between VSTM accuracy and maintenance time (β = 0.018, z = 2.171, p = 0.030, Fig. 3a), indicating that the ripple ramping-up effects differed between remembered and forgotten trials. Post hoc analyses revealed a significant ramping-up for remembered trials (β = 0.014, z = 5.024, p_FDR < 0.001), but not for the forgotten items (β = −0.002, z = −0.282, p_FDR = 0.778).

Figure 3.

Further control analyses revealed that the VSTM accuracy-related hippocampal ramping-up effect was robust across alternative ripple definitions (e.g., ripple duration ≥ 25 ms or alternative frequency bands: 80-120 Hz and thresholds following previous research¹⁸; see also STAR Methods and Fig. S3) and remained significant using conventional group-level analysis (STAR Methods and Fig. S3; see also Table S1 for participant-level ramping-up slopes). Moreover, the hippocampal ramping-up effect was unrelated to retrieval speed (β = 0.009, z = 1.570, p = 0.117; Fig. S3) and did not differ between novel and repeated trials (β = −0.009, z = −1.716, p = 0.086; Fig. S3).

We also tested the ripple ramping up during maintenance in the LTL. Unlike the hippocampus, no significant ramping-up was observed (β = −0.005, z = −1.772, p = 0.076, see Fig. S3), nor was the change in ripple rates over maintenance associated with VSTM accuracy (time ×accuracy interaction effect: β = −0.010, z = −1.246, p = 0.213; remember: β = −0.006, z = −2.033, p_FDR = 0.084; forget: β = 0.004, z = 0.507, p_FDR = 0.612; Fig. 3b) or RT (time ×RT interaction effect: β = 0.006, z = 1.087, p = 0.277, see also Fig. S3). To ensure that the absence of a ramping-up effect was not due to low signal-to-noise channels, we restricted the analysis to 89 bipolar channel pairs with at least one contact in LTL gray matter or both contacts within 2 mm of it, and the results remained (see Fig. S3). In addition, the absence of the ramping-up effect in the LTL cannot be attributed to greater signal heterogeneity compared with the hippocampus (see Fig. S3).

Hippocampal-LTL coupled ripples associate with VSTM performance

Previous research suggested that hippocampal ripples support memory via hippocampal-neocortical interactions³⁵, which may be supported by ripple couplings across regions^18,37,42. We therefore examined whether hippocampal-LTL ripple coupling supports VSTM. Following previous studies^43,44, we computed the LTL ripple rate within ±0.5 s of each hippocampal ripple peak (Fig. 4a; See STAR Methods) across all task stages as well as within each task stage. For comparison, we computed LTL ripple rates within ±0.5 s of randomly selected, hippocampal ripple-free time points for 1000 times, resulting in a surrogate distribution. The empirical coupled ripple rates were then z-scored against the surrogate distribution to yield a normalized coupled ripple index (see STAR Methods). Our results revealed that the normalized hippocampal-LTL coupled ripple index was significantly greater than zero across all task stages (p_cluster < 0.011; Fig. 4b), and within individual stages (all ps_cluster < 0.047).

Figure 4.

To further investigate whether coupled ripples contribute to VSTM performance, we compared coupled ripple rates between remembered and forgotten trials. The results revealed significantly higher coupled ripple rates in remembered trials compared to forgotten trials during maintenance (ps_cluster < 0.019; Fig. 4c). This effect remained significant after matching trial counts between remember and forget conditions using bootstrapping (ps < 0.040). No significant differences were observed during encoding or retrieval stages (ps_cluster > 0.189). Together, these findings suggest that hippocampal-LTL ripple coupling during the maintenance associated with successful VSTM.

Hippocampal-LTL coupled ripples coordinate memory reactivation in the LTL

We next tested whether hippocampal-LTL coupled ripples support the VSTM reactivation in the LTL. To this end, we applied a multivariate decoding method to the spectral power of LTL channels for individual participants to classify categories of learned pictures (see STAR Methods). Classifiers were trained on each encoding time window and tested across encoding, maintenance, and retrieval time windows. Among remembered trials, we identified significant clusters showing decoding accuracies above chance level (25%) during encoding, early maintenance, and retrieval (all ps_cluster < 0.039, Fig. 5a). When averaged across all time windows within each stage, decoding accuracy remained significantly above chance for individual stages (all ps_FDR < 0.003, Fig. 5b). Note that the above-chance decoding during the maintenance stage cannot be simply attributed to the lingering sensory input immediately following encoding. This effect remains robust in the 3-4 s post-encoding interval (Fig. S4). Furthermore, remembered items showed greater decoding accuracy than forgotten ones during the pre-response retrieval stage (Fig. S4). These results suggested that VSTM is represented and reactivated in the LTL.

Figure 5.

We then tested whether this reactivation was temporally locked to the hippocampal-LTL coupled ripples during remembered trials. To enhance signal-to-noise ratio, we pooled hippocampal-LTL coupled ripples across all task stages (mean ±SD: 439 ±125 coupled ripples per participant) and aligned decoding accuracy to ±0.5 s around hippocampal ripple peaks of the LTL coupled ripples. Notably, the decoding accuracy was z-scored relative to the surrogate distribution, where decoding accuracy aligned to non-coupled ripple time points, and then tested against zero (see STAR Methods). Our results revealed a significant positive cluster when the classifier was trained on late encoding windows (1.85-2.25 s post-stimulus; p_cluster = 0.029; Fig. 5c), indicating LTL memory reactivation time-locked to coupled ripples. Further analysis based on the identified cluster, we found that this coupled ripple-locked decoding accuracy was significant during maintenance (p_cluster = 0.004, Fig. 5d), but not during the encoding or retrieval stage (ps_cluster > 0.118). Notably, coupled ripple-locked LTL decoding accuracy in the clusters in Fig. 5c and 5d was significantly above chance (i.e., 25%; all ps < 0.002).

As a control analysis, we also examined memory reactivation in the hippocampus and its temporal link with hippocampal-LTL coupled ripples. Although hippocampal decoding accuracy was above chance when averaged across time windows for each task stage (ps_FDR < 0.031), it was not significantly time-locked to hippocampal-LTL coupled ripples (ps_cluster > 0.470; Fig. S5). Further control analyses to test the specificity of coupled ripple revealed that decoding accuracy in neither the hippocampus nor the LTL was modulated by independent, uncoupled ripples (i.e., hippocampus or LTL ripples without coupling) (ps_cluster > 0.203; Fig. S5). These findings suggest that coupled hippocampal-LTL ripples coordinate memory reactivation in the LTL, rather than in the hippocampus, during VSTM maintenance.

VSTM-related ripple dynamics are not associated with subsequent long-term memory performance

To rule out potential confounds from long-term memory (LTM) formation, we separated VSTM-remembered trials based on whether they were later remembered or forgotten in the LTM test and compared the hippocampal ramping-up effect, hippocampal-LTL coupled ripples, and coupled ripple-locked memory reactivation between these two conditions. Our results showed that the hippocampal ripple ramping-up effect was significant for both subsequently LTM remembered (β = 0.017, z = 4.141, p_FDR < 0.001) and forgotten trials (β = 0.011, z = 2.987, p_FDR = 0.003; see Fig. S6). Critically, no significant difference was observed between subsequently LTM remembered and forgotten items (β = 0.003, z = 0.439, p = 0.661). Moreover, neither the coupled ripple rate nor ripple-locked reactivation differed between subsequently LTM remembered and forgotten trials (Fig. S6). Together, these control analyses suggest that our main findings reflect mechanisms underlying VSTM rather than LTM.

Discussion

While previous rodent and recent human studies have implicated hippocampal ripples in long-term memory consolidation and retrieval^16,45, it remains unclear whether and how hippocampal ripples support human VSTM. Our study provides novel electrophysiological evidence that hippocampal ripples and their coordination with the neocortex support VSTM. We found that hippocampal ripples progressively ramped up during the maintenance period, and this ramping-up effect was associated with successful VSTM. Moreover, hippocampal ripples were temporally coupled with LTL ripples, and these coupled events coincided with memory reactivation in the LTL during maintenance.

First of all, no overall above-baseline hippocampal ripple rate when averaged across the 7-s maintenance time windows, combined with the above-baseline ripple rates during late maintenance and the ripple ramping-up effect over the 7-s maintenance period, collectively aligns with the dynamic coding framework of VSTM³⁰. The ramping-up in hippocampal ripple rates is associated with memory accuracy but not reaction time, suggesting its role in proactive memory retrieval during VSTM rather than general arousal or motor preparation^46,47. Moreover, this ramping-up effect is unlikely to reflect signal drifts over long intervals, since it was observed exclusively in remembered trials and has also been reported in studies using shorter retention periods, such as 1.5 s⁴⁸ or 3 s⁴⁹. Nonetheless, we cannot rule out the possibility that the magnitude of this effect may be affected by the duration of the delay period. Future studies could systematically manipulate retention intervals to test this possibility. In addition, we found hippocampal ripple rates decreased immediately following the response, a pattern consistent with that during episodic memory retrieval^17,40, which may reflect diminished hippocampal engagement following memory retrieval.

While our findings align with the dynamic coding framework of VSTM, they appear to contrast with prior reports of persistent hippocampal spiking during short-term memory maintenance^10,11. Several factors may account for this discrepancy, including differences in neural measures, task demands. First, prior studies identifying persistent activity were mostly based on single-unit recordings^10,11,23, revealing sustained firing in a minority of stimulus-selective neurons. In contrast, our local field potential (LFP) recordings may reflect a dynamic coding scheme at the population level³⁰. Compared to persistent firing, this dynamic population coding is less constrained by memory capacity limits and may offer more flexible control while minimizing inter-item competition⁵⁰. Second, contemporary short-term memory models suggest that items within the focus of attention are maintained via persistent activity, while unattended items are stored in an activity-silent state^51,52. Therefore, multi-item VSTM tasks with higher attentional demands may favor persistent activity, while our single-item task may have allowed memory to drift out of focus, favoring an overall activity-silent state with discrete reactivation bursts. Corroborating this possibility, a previous study found persistently increased hippocampal gamma band activity during multi-item working memory maintenance but not for the single-item²⁶.

Critically, our findings suggest that hippocampal ripples serve as high-frequency bursts that coordinate memory representational refresh during the short-term memory maintenance period. While these ripples are sparse in time, they fit the proposed functional role of high-frequency bursts that intermittently refresh synaptic weights, thereby maintaining the learning-induced, short-lived changes^28–30. Corroborating our finding, recent rodent research shows that hippocampal activity during VSTM delays is characterized by low-rate activity that supports memory reactivation⁵³. Complementing these studies, we show that hippocampal ripples coordinate memory reactivation through coupling with LTL ripples, suggesting that hippocampal ripples refresh memory traces not only in the hippocampus but also in the neocortex. Together with the previous human intracranial studies of hippocampal ripples during episodic memory^16,17,40, these findings convergently support the functional significance of low-rate hippocampal activity.

In addition, we emphasize that hippocampal ripples are unlikely to be the sole mechanism supporting VSTM. For example, prior work has shown that gamma bursts in the prefrontal cortex gate access to encoded VSTM representations and that ramping of such bursts supports working memory in primates⁴⁸. Moreover, recent rodent research demonstrates that replay sequences can occur even in the absence of sharp-wave ripples⁵⁴. These findings suggest that ripple ramping-up should be interpreted as one component within a broader, distributed set of dynamic mechanisms supporting short-term memory maintenance. However, the use of iEEG in humans limited us from direct observation of full brain network, and future studies combining broader spatial coverage with high temporal resolution will be essential for characterizing how hippocampal ripples interact with complementary maintenance mechanisms.

Notably, our findings show that memory reactivation was locked to coupled ripples in the hippocampus and LTL, rather than isolated ripples in either the hippocampus or the LTL, highlighting the necessity of hippocampal-neocortical communication. Our work further demonstrated that coupling strength was associated with VSTM accuracy, which is consistent with prior findings showing that increased hippocampal-neocortical functional connectivity, from pre- to post-encoding rest, is associated with subsequent memory performance⁵⁵. Comparing with previous research emphasizing low-frequency coordination supporting VSTM —such as theta-band (4-5 Hz) phase coherence between hippocampus and neocortex⁵⁶ and 1-10 Hz oscillations modulating memory reactivation⁵⁷, our findings suggest hippocampal-cortical communication could be achieved through an alternative mechanism: the co-occurrence of ripples. Consistent with our findings, the latest evidence shows that ripples in different brain areas co-occur with near-zero phase lag, providing a temporally precise mechanism for efficient inter-regional communication⁴². These different neural mechanisms—low-frequency coherence and ripple co-occurrence—may not necessarily be mutually exclusive. One possibility is that hippocampal ripples are nested within ongoing theta oscillations during cross-regional communication⁵⁸, a hypothesis that remains to be tested in future research.

We acknowledge that the use of familiar, semantically rich everyday objects in our delayed match-to-sample task likely engages long-term memory (LTM) processes. Prior work suggested that short-term performance can be supported by pre-existing semantic knowledge and episodic experience, such that hippocampal engagement may reflect interactions between visual perceptual input and long-term pre-existing knowledge/episodes, rather than a strictly process-pure visual short-term memory (VSTM)^59–61. Although prior work has often used abstract stimuli to minimize episodic contributions⁶², disentangling VSTM from LTM was not the primary aim of this study. Moreover, recent evidence suggests that simplified stimuli may underestimate how the brain maintains meaningful information in real-world contexts⁶¹. Indeed, semantically rich stimuli can enhance working memory capacity by enabling access to long-term representations^63–66. We therefore interpret our findings as reflecting mechanisms that support the short-term maintenance of meaningful, real-world visual representations. Future studies aiming to more strictly dissociate VSTM from LTM contributions should employ artificial or abstract stimuli that minimize pre-existing knowledge associations.

Moreover, our findings showed that both the ramping-up hippocampal ripple rates and the hippocampal-LTL coupled ripples during maintenance were associated with VSTM rather than subsequent LTM accuracy. These findings suggest that the ripple dynamics we observed primarily support the short-term maintenance of complex visual features and their binding within the pictures themselves rather than the formation of long-term memory for cue-picture pairs, which has been shown in previous research³⁹. This interpretation is strengthened by recent studies showing that hippocampal ripples and hippocampal neuronal firing support VSTM of complex naturalistic pictures in the absence of subsequent LTM tests or cue-picture associations^67,68. Moreover, converging evidence from studies on neural oscillations, patients with focal hippocampal damage, and neural modulation demonstrates that the hippocampus plays a critical role in VSTM, where high-resolution feature binding is required^4,8,9.

Beyond binding demands, the hippocampus also contributes to VSTM for simple visual features that minimally engage LTM, such as color squares, when memory precision is dissociated from binding errors or when relational demands are minimized^4,69. For instance, previous research found that patients with hippocampal damage exhibited reduced precision for simple color memories after brief delays, yet showed no increase in relational binding errors⁴. Similarly, a recent study demonstrated that MTL lesions selectively impair the precision of VSTM representations rather than the quantity of items retained⁶⁹. In addition, the hippocampus is also known to engage in tasks with relatively low memory precision demands, such as change detection paradigms requiring only coarse-level discrimination between targets and distinct lures^6,10,67. These findings collectively indicate that the hippocampus supports a wide range of VSTM tasks. Future work should systematically examine whether key factors, such as relational binding demands and memory precision, shape hippocampal ripple dynamics, using experimental manipulations or multi-component modeling⁷⁰.

The absence of differences in ripple rates and category-level decoding accuracy between VSTM remembered and forgotten trials during maintenance may reflect that hippocampal ripples serve dual functions. While they support high-fidelity memory reactivation to discriminate the highly similar lures from target pictures in remembered trials, they may also coordinate coarser, less precise reactivation for forgotten trials. These coarse representations are likely insufficient and less accessible to support fine discrimination. Supporting this interpretation, we observed that memory decoding evidence was lower for forgotten items compared with remembered items prior to VSTM retrieval responses.

In addition, the ripple ramping-up effect was only observed in the hippocampus but not in the LTL, despite the well-documented role of LTL in representing VSTM content^39,71. The absence of LTL ripple ramping-up cannot be attributed to more heterogeneous ripple activity in LTL channels. However, our LTL coverage spanned inferior, middle, and superior temporal lobes, due to the patients’ clinical implantation scheme, which prevented us from restricting analyses to specific subregions. Consequently, we cannot rule out the possibility that ramping-up effects may exist in more restricted LTL subregions. In addition, our results raise an important question of whether similar ripple ramping-up effects occur in other brain regions. Future studies with broader and more targeted electrode coverage, including prefrontal and occipital cortices known to interact with the hippocampus during VSTM, will be necessary to address these questions^72,73.

To conclude, our study provides direct electrophysiological evidence that hippocampal ripples—and their coordination with neocortical regions—support VSTM. We show that ripple activity in the hippocampus ramps up during maintenance and is associated with memory accuracy. These findings support a dynamic coding model of VSTM and suggest that hippocampal ripples orchestrate discrete reactivations that sustain latent representations. By linking ripple dynamics to both representational reactivation and interregional coupling, our results extend the hippocampus’s role from long-term formation to short-term memory maintenance and offer new insights into the neural mechanisms unifying these memory systems.

STAR★Methods

Key resources table

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact: Ying Cai (yingcai@zju.edu.cn).

Materials availability

This study did not generate new unique reagents.

Experimental model and study participant details

Participants

We reanalyzed the data from a previous study⁵⁷. We included thirteen epilepsy patients (mean age ±SD: 26.77 ±5.48 years, 7 females) who had electrodes implanted in both the hippocampus and LTL. The iEEG data were recorded at the Center of Epileptology, Xuanwu Hospital, Capital Medical University, Beijing, China. The study adhered to the latest version of the Declaration of Helsinki and was approved by the Institutional Review Board at Xuanwu Hospital. All participants have a normal or corrected-to-normal visual acuity. They all signed written informed consent before the experiment.

Method details

Stimuli and procedures

The study consisted of a visual short-term memory (VSTM) task, followed by a 1-minute countback task and a short break (1 to 4 minutes), and then a cued-recall long-term memory test. The study used 56 pictures of familiar everyday objects and 112 two-character Chinese verbs. The pictures were drawn from four categories—fruits, animals, electrical devices, and furniture—with 14 images per category. Each picture was randomly paired with two different cue words across two consecutive experimental runs.

The VSTM was assessed using a modified Delayed Match-to-Sample (DMS) task. In each run, participants studied 14 unique word-picture pairs, each repeated three times, resulting in 42 trials per run. Each trial started with a brief fixation period (300 ms), followed by a jittered blank interval (800-1200 ms). A cue word and its associated target picture were then presented centrally for 3 s, during which participants were instructed to encode the association between the word and the picture. Immediately following the encoding stage, the picture disappeared while the cue word remained on screen for 7 s. During this maintenance stage, participants were instructed to mentally maintain a vivid image of the target picture. Displaying the cue word during maintenance served to reinforce the word-picture association, reduce working memory load, and minimize distraction. It was then followed by an immediate retrieval stage, during which a probe picture was displayed on the screen, and participants were instructed to determine whether it matched the target picture by pressing one of two buttons within 2 s. The probe picture was either the same as the target picture (50%, match trials) or a highly similar lure picture (50%, nonmatch trials). Each participant completed between 4 and 8 DMS runs (mean ±SD: 6.14 ±1.46), resulting in a total of 168-336 trials (mean ±SD: 258.00 ±61.32) per participant.

In the cued-recall long-term memory test, a cue word was presented on screen, and participants reported the category of the associated picture by pressing one of four buttons corresponding to the four object categories. The current study focused primarily on VSTM stages, with the cued-recall data included as a control to help dissociate VSTM-related processes from long-term memory.

Quantification and statistical analysis

Data recordings and preprocessing

Intracranial EEG data were recorded from depth electrodes, each containing 12 or 16 channels (2 mm in length, 0.8 mm in diameter, spaced 1.5 mm apart). Data were collected using amplifiers from Brain Products (Brain Products GmbH), NeuroScan (Compumedics Limited), or Nicolet (Alliance Biomedica Pvt Ltd.) electroencephalogram systems, with the sampling rates of 2,500, 2,000, and 2,048 Hz, respectively. During the online recordings, data from all channels were referenced to a common subcutaneous channel. During the offline preprocessing, we first removed the channels within the epileptic loci and channels that were severely contaminated by epileptic activity. The remaining channels in the hippocampus (HPC) and lateral temporal lobe (LTL) were visually inspected and bipolar referenced to channels on the same electrode. Because not all participants had multiple hippocampal channels on the same electrode, HPC channels were bipolar re-referenced to the nearest white matter channel^40,74. LTL channels were bipolar re-referenced to an adjacent channel within the LTL³⁹.

Channel localization

The identification of channel locations included the following steps. First, each participant’s post-implantation CT scan was co-registered with their pre-implantation MRI. The MRI was then normalized to Montreal Neurological Institute (MNI) space using Statistical Parametric Mapping 12 (SPM12). Anatomical localization and 3D visualization of electrode channels were conducted using the 3D Slicer platform (https://www.slicer.org/). To assign anatomical labels, structural MRIs were segmented with FreeSurfer (https://surfer.nmr.mgh.harvard.edu), and the nearest cortical or subcortical label was assigned to each channel. Channels in the hippocampus were further verified through visual inspection in each participant’s native anatomical space.

Ripple detection and peri-ripple spectral analysis

Ripple detection followed the established procedures outlined in previous studies^16,17,40. First, a 70-180 Hz bandpass linear-phase Hamming-windowed finite impulse response (FIR) filter was applied to the bipolar re-referenced iEEG data on HPC and LTL channels, with a transition width of 5 Hz. We then computed the analytic signal amplitude (squared) using the Hilbert transform. To determine the ripple detection threshold, we: (1) identified and clipped extreme values (≥4 SD above mean) using the Least-Median-Square (LMS) method to reduce outlier bias; (2) smoothed the clipped signal with a 40 Hz low-pass Kaiser-windowed FIR filter (5 Hz transition); (3) calculated the mean and SD of ripple-band amplitude across all runs per channel. Candidate ripples were defined as periods when the squared amplitude exceeded 4 SD above the mean, with start/end points marked by crossings of 2 SD. Events lasting < 20 ms or > 200 ms were excluded. Ripple peaks were identified with MATLAB’s findpeaks.m, and events with < 30 ms peak-to-peak intervals were merged.

To validate the results from the above ripple detection approach, we have also performed an alternative ripple detection method following previous research¹⁸. Specifically, we first bandpass-filtered the iEEG signal in the ripple range (80-120 Hz) using a second-order Butterworth filter and then applied a Hilbert transform to extract the instantaneous amplitude. Candidate events were identified when the Hilbert envelope exceeded 2 SD above the mean amplitude. Events were kept as ripples only if they lasted at least 25 ms and reached a peak amplitude >3 SD. Adjacent events separated by <15 ms were merged.

To prevent contamination from pathological activity, interictal epileptiform discharges (IEDs) were rigorously screened. Channels within clinically identified epileptogenic zones were excluded. For the remaining channels, IEDs were independently verified by two neurologists at Beijing Xuanwu Hospital. IEDs typically exhibit broadband power increases (1-180 Hz), whereas physiological ripples show narrow high-frequency activity (70-180 Hz). Candidate ripple events exhibiting broadband spectral increases or occurring within 50 ms of an IED were excluded.

To obtain the peri-ripple spectrograms, we first epoched iEEG data into 6-s segments, ranging from 3 s before the ripple peak to 3 s after it. The epoched data were then convolved with complex Morlet wavelets (six cycles) in the range of 1 to 250 Hz, with 2 Hz steps. The resulting complex wavelet transform was squared to obtain spectral power. Power values were z-transformed for each frequency and channel using the mean and the SD of the power across all epochs belonging to the same experimental run. Final spectrograms were centered around ripple peaks (±250 ms) in 500-ms windows. All iEEG analyses were performed in MATLAB using custom scripts and functions from the FieldTrip toolbox⁷⁵.

Analysis of ripple rate changes and cluster-based permutation test

Detected ripples were time-locked to stimulus onset for the encoding and maintenance stages, and to behavioral responses for the retrieval stage. For each trial, ripple rate was calculated by dividing the number of ripples by the duration (in seconds) of each task stage: encoding (0-3 s post-stimulus onset), maintenance (3-10 s), and retrieval (from probe onset to response). The ripple rate in a pre-encoding time window (200 to 800 ms prior to stimulus onset) was also computed and served as the baseline.

We used linear mixed-effect model (LMM) analysis to compare the ripple rates between each VSTM task stage and the pre-encoding baseline, as well as between different task stages. The analysis was performed using the R package lme4⁷⁶. The models were specified as follows:

Here, 𝑟𝑖𝑝𝑝𝑙𝑒_𝑟𝑎𝑡𝑒 refers to the ripple rate per channel, and 𝑠𝑡𝑎𝑔𝑒𝑠 is a fixed effect with four levels: pre-encoding baseline, encoding, maintenance, and retrieval. Random intercepts were included for each participant and for trials nested within participants to account for within-subject and within-trial variability. Pairwise comparisons between stages were performed using the emmeans package, and the resulting p-values were corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) procedure⁷⁷.

To quantify ripple rate changes at a finer temporal resolution, we computed ripple rates using non-overlapping 100-ms time windows within each task stage. The resulting time series was smoothed using a 400-ms Gaussian kernel. Ripple rates were then compared to the pre-encoding baseline separately for the encoding, maintenance, and retrieval stages. To correct for multiple comparisons, we performed non-parametric cluster-based permutation tests using the MATLAB codes (https://doi.org/10.5281/zenodo.10877825). Specifically, for each time point, we computed the empirical ripple rate difference between conditions (e.g., encoding vs. baseline). We then compare it to a null distribution of ripple rate differences, which were generated by randomly permuted condition labels (e.g., encoding vs. baseline) 5,000 times. At each iteration, the ripple rate difference between conditions was recalculated. The empirical threshold for significance (α = 0.05, two-tailed) was determined from this null distribution. Time points of empirical ripple rate difference exceeding this threshold were grouped into clusters. For each cluster, a cluster-level statistic was computed by summing the ripple rate differences within the cluster. The observed cluster statistics were then compared against the distribution of maximum and minimum cluster-level statistics derived from the 5,000 permutations to determine significance. This same cluster-based permutation procedure was also used to compare ripple rates between remembered and forgotten trials, and between fast and slow remembered trials.

Ramping-up effects analysis

To test whether the ramping-up of ripple activity during maintenance predicted VSTM performance, we fit generalized linear mixed-effects models (GLMMs) with a Poisson link function. The main model was specified as:

Here, ripple refers to the trial-level ripple count (Poisson-distributed). Time corresponds to the 7-s maintenance time index, and memory reflects VSTM accuracy (remembered = 1, forgotten = 0) or response speed (fast vs. slow), entered as fixed factors. Random intercepts and time slopes were included for each participant, and random intercepts for trials nested within participants. When this time ×memory interaction effect was significant, we conducted follow-up analyses by fitting separate GLMMs for remembered and forgotten trials (or fast and slow trials). These post-hoc models tested whether ripple activity increased over time within each memory outcome, and corresponding p-values were further FDR corrected.

To validate our GLMM results, we also performed a group-level ramping-up analysis. Ripple rates were first averaged across trials and channels, yielding a time series of mean ripple rates during the maintenance stage for each participant. A linear regression was then applied to each participant’s ripple rate time series, with the regression slope quantifying the change in ripple rates over time. Independent t-tests were used to assess whether slopes were significantly greater than zero for remembered and forgotten trials, and to compare slopes between these two conditions.

Coupled ripple analysis

To identify the coupling between hippocampal and LTL ripples, we tested whether the LTL ripple rates in the presence of hippocampal ripples were larger than in the absence of hippocampal ripples^43,44. The hippocampus channels were paired with all LTL channels in each participant. For each hippocampal-LTL channel pair, we computed the hippocampal ripple peak-locked LTL ripple rate in a ([-0.5 s, 0.5 s]) time window. Additionally, non-ripple surrogates were derived for each channel pair, and the surrogate-locked LTL ripple rates were computed in time windows of the same length. Specifically, for all the n hippocampal ripples, we randomly selected n time points in the hippocampal channel with no ripple occurring 0.5 s before and after. Then, the LTL ripples were aligned to these non-ripple surrogates to obtain the surrogate-locked LTL ripple rates. This procedure was repeated 1000 times to generate a surrogate distribution, and empirical hippocampal ripple-locked LTL rates were z-scored relative to it, resulting in the normalized hippocampal-LTL coupled ripple indexes. The normalized hippocampal-LTL coupled ripple indexes were then averaged across all hippocampal-LTL channel pairs for each participant. Cluster-based permutation tests assessed whether normalized hippocampal-LTL coupling indexes exceeded zero within the ± 0.5 s window across participants. Significantly above-zero coupling index indicates hippocampal-LTL coupled ripples. This analysis was conducted across all task stages and within each of the task stages (i.e., encoding, maintenance, and retrieval).

Multi-variate decoding of VSTM representations

To decode category-specific memory representations, we trained a linear support vector machine (SVM) to classify the four picture categories (animal, fruit, electrical device, furniture) from remembered and forgotten trials. The chance level for decoding was set at 25%. Inspired by prior studies showing that memory content is best captured by a broad range of spectral power^78,79, we extracted broadband (2-180 Hz) spectral power from hippocampal and LTL channels. Time-frequency transformation was performed using complex Morlet wavelets (2-29 Hz in 1-Hz steps and 30-180 Hz in 5-Hz steps, 6 cycles), and all power spectral data were down-sampled to 100 Hz after time-frequency transformation. Spectral power was z-scored for each frequency and channel across runs. To obtain time-resolved memory reactivation, decoding was performed in sliding time windows for both the training and test data, with a 400-ms window length and a 50-ms increment.

To increase the signal-to-noise ratio, the spectral power was averaged across time points within each time window, resulting in frequency ×channel features as input to the SVM model. For each participant, normalized power from all hippocampus or LTL channels was used as input to the SVM. A leave-one-trial-out cross-validation was used to estimate decoding accuracy. SVMs were trained separately on each time window during encoding (yielding 60 decoders) and tested across all time windows of the task to assess temporal generalization. Decoding accuracy at each time point was statistically compared against the 25% chance level using the cluster-based permutation test.

Coupled ripple-locked decoding accuracy

To assess whether memory reinstatement was linked to coupled ripples between the hippocampus and LTL, we computed decoding accuracy time-locked to coupled ripple events. First, coupled ripple events were identified separately for each hippocampal (HPC)-lateral temporal lobe (LTL) channel pair on individual trials. Coupled ripples were defined as temporally overlapping hippocampus and LTL ripples or those occurring within 50 ms of each other¹⁸. The peak time of each coupled event was set to the hippocampal ripple peak. The coupled ripple counts were then pooled across all trials and channel pairs within each participant, and subsequently averaged across participants. On average, each participant exhibited 439 (SD:125) coupled ripples in the VSTM task. When broken down by stage, each participant showed 143 ±45 coupled ripples during encoding, 274 ±76 during maintenance, and 23 ±6 during retrieval.

We then aligned trial-level decoding accuracy to these coupled ripple peaks, averaging values within a ±0.5 s window to obtain the coupled-ripple-locked decoding accuracy. In addition to using the chancel level (25%) as a baseline, we also generated decoding accuracy locked to non-coupled surrogate events as the alternative baseline. Specifically, for each of the coupled ripples, a corresponding time point without any coupled ripple activity within ±0.5 s was randomly selected. This process was repeated 1000 times to produce a surrogate distribution. Empirical coupled ripple-locked decoding accuracy was z-scored relative to this distribution to obtain normalized decoding accuracy. Cluster-based permutation tests were used to assess whether normalized decoding accuracy exceeded zero.

As a control, we examined decoding accuracy locked to independent hippocampal or LTL ripples—i.e., ripples not overlapping or occurring within 50 ms of a ripple in the other region. Decoding accuracy for these independent ripples was z-scored relative to a non-ripple surrogate distribution, and the same statistical procedures were applied.

Supplementary information

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Table S1.

Data availability

All preprocessed data are available on the Open Science Framework: https://osf.io/gwe62/. All custom MATLAB code necessary to reproduce the main conclusions of this study are available on the Open Science Framework: https://osf.io/gwe62/2. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgements

We thank the patients who volunteered to participate in the experiments and the support from our collaborators from Beijing Xuanwu Hospital, who provided the patient resources. This work was supported by the STI 2030-Major Projects (2021ZD0200401, 2021ZD0200409), the Fundamental Research Funds for the Central Universities (226-2024-00118), National Natural Science Foundation of China (32100851), and Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515012667 and No. 2023A1515110311). This work was also supported by the grant from the Research Center for Brain Cognition and Human Development, Guangdong, China (No. 2024B0303390003).

Additional information

Author contributions

J.L. and G.X. conceived the original experiment. J.L., X.H., C.Y., N.A., S.Z., and Y.C. performed the analysis. J.L., X.H., and Y.C. wrote the initial manuscript. J.L., X.H., C.Y., N.A., G.X., S.Z., and Y.C. revised the manuscript.

Funding

STI 2020-Major Projects (2021ZD0200401)

• Ying Cai

STI 2020-Major Projects (2021ZD0200409)

• Ying Cai

The Fundamental Research Funds for the Central Universities (226-2024-00118)

• Ying Cai

MOST | National Natural Science Foundation of China (NSFC) (32100851)

• Ying Cai

Guangdong Basic and Applied Basic Research Foundation (2024A1515012667)

• Jing Liu

Guangdong Basic and Applied Basic Research Foundation (2023A1515110311)

• Jing Liu

The Research Center for Brain Cognition and Human Development (2024B0303390003)

• Jing Liu