Figures and data
Experimental paradigm, stimuli, and intracranial EEG channel localization.
(a) Upper: an exemplar trial procedure of the delayed match-to-sample (DMS) task. Lower: examples of target and lure pictures. Pictures were selected from four categories—animals, fruits, electrical appliances, and furniture—with each row representing one category; (b) Normalized locations of intracranial EEG channels across all 13 participants. Pink spheres indicate hippocampal (HPC) channels within the highlighted yellow-shaded region, hippocampus; blue spheres indicate lateral temporal lobe (LTL) channels.
VSTM task-induced ripple rate dynamics.
(a) Examples of ripple activities from one hippocampal (HPC) channel. Upper (from left to right): a peri-ripple raw iEEG; ripple bandpass (70-180 Hz) filtered iEEG; Power spectrum of the peri-ripple iEEG. Lower (from left to right): averaged peri-ripple raw iEEG across ripples in a channel; averaged ripple bandpass filtered iEEG; averaged peri-ripple power spectrum; (b) Upper: Ripple raster plot for all individual trials from an example HPC channel. Trials are sorted according to reaction times. Each pink dot represents the peak time of a ripple. Lower: Time-resolved hippocampal ripple rates averaged across participants and channels during encoding, maintenance, and retrieval. The shaded areas around the lines indicate the standard error of the mean (SEM). The three vertical lines from left to right indicate onsets of encoding, maintenance, and retrieval response, respectively. The bolded grey curve represents the probe onset of each trial during retrieval; Black horizontal bars on the top of the lower panel indicate time clusters where ripple rates differ significantly from the pre-encoding baseline (pcluster < 0.05). (c) Linear mixed-effect model (LMM) estimated hippocampal ripple rates for pre-encoding baseline (0.2 to 0.8 s before stimulus) and individual task stages. Encoding: 0-3 s; maintenance: 3-10 s; retrieval: probe onset to response. (d) Upper: Ripple raster plot from an example LTL channel; Lower: Time-resolved LTL ripple rates averaged across participants and channels during encoding, maintenance, and retrieval. (e) LMM estimated LTL ripple rates for baseline and individual task stages. Black asterisks indicate significantly higher ripple rates relative to the pre-encoding baseline or significant differences between task stages. *: pFDR < 0.05, **: pFDR < 0.01, ***: pFDR < 0.001, ns: not significant.
Ripple ramping-up effects during maintenance.
(a) Hippocampal (HPC) ripple ramping-up effects for remembered vs. forgotten trials. (b) Lateral temporal lobe (LTL) ripple ramping-up effects for remembered vs. forgotten trials. The shaded areas around the lines indicate the SEM. β: estimated fixed effect coefficients for remember or forget conditions. ***: pFDR < 0.001.
Hippocampal-LTL coupled ripples.
(a) Left: Illustration of coupled ripples between HPC and LTL (first shaded area) and uncoupled HPC ripple (second shaded area) and LTL ripple (third shaded area). Middle: LTL ripple rates time-locked to an exemplar HPC ripple from one participant; Right: LTL ripple rates time-locked to surrogate time points without HPC ripples from the same channels. Each row indicates LTL ripples locked to a single HPC ripple peak or surrogate time point. Each blue dot represents an LTL ripple, and the curve shows LTL ripple rates across all trials surrounding HPC ripple peaks or surrogate time points. (b) Normalized hippocampal-LTL coupled ripple index across all task stages (i.e., task average) and within individual task stages. (c) Hippocampal-LTL coupled ripple rate difference in remembered versus forgotten trials. Black bars at the top indicate time windows with significant differences between conditions (survived after cluster-based permutation tests: pcluster < 0.05). The shaded areas around the lines indicate the SEM.
Coupled ripples coordinate memory reactivation in the LTL.
(a) LTL decoding accuracy of remembered trials compared to chance level (25%) across the task (left: encoding and maintenance, 0 indicates stimulus onset; right: retrieval, 0 indicates behavior response). Clusters with significantly above-chance decoding accuracy (survived cluster-based permutation test) are circled by black lines. (b) Decoding accuracies during encoding, maintenance, and retrieval stages are significantly above chance. (c) Normalized decoding accuracy time-locked to hippocampal-LTL coupled ripples relative to surrogate distribution. The black-circled cluster indicates normalized decoding accuracy significantly above zero. (d) Coupled ripple-locked normalized decoding accuracy averaged across the late encoding cluster identified in (c). Black bars at the top indicate significant clusters with normalized decoding accuracy significantly above zero. All clusters survived cluster-based permutation tests (pcluster < 0.05). The shaded areas around the lines indicate the SEM. **: pFDR < 0.01.
Properties of ripples in HPC and LTL channels.
Distributions of ripple rate, peak frequency, duration, and amplitude across all HPC channels.
Ripple rates in the hippocampus and lateral temporal lobe (LTL) during the visual short-term memory (VSTM) task and long-term memory (LTM) retrieval.
(a) Time course of hippocampal ripple rates for novel (first presentation) and repeated (second and third presentations) trials. The black horizontal bar indicates a significant cluster (500–1100 ms post-stimulus) with higher ripple rates for novel versus repeated trials (pcluster < 0.001). (b) No significant differences were observed between hippocampal ripple rates for remembered and forgotten trials in any VSTM stage (all pscluster > 0.466). (c) Hippocampal ripple rates were significantly higher for LTM remembered than forgotten trials within a cluster (i.e., 1600-1900 ms before LTM retrieval responses) for remembered than forgotten trials (pcluster = 0.047). (d) Hippocampal ripple rates were significantly higher for fast compared with slow trials within a pre-response cluster (i.e., 150-450 ms before retrieval responses, pcluster = 0.014). (e) No significant differences were observed between LTL ripple rates of remembered versus forgotten trials in any VSTM stage (all pscluster > 0.274). (f) No significant differences were observed between hippocampal ripple rates for fast and slow remembered trials in any VSTM stage (all pscluster > 0.237). Shaded areas represent ±SEM.
Control analyses for the ripple ramping-up effects in the hippocampus (HPC) and lateral temporal lobe (LTL).
(a) Significant ramping-up of hippocampal ripple rates during the maintenance period across all trials, channels, and participants (β = 0.012, z = 4.678, p < 0.001). The y-axis shows model-estimated ripple rates. (b) Group-level analysis of the ripple ramping-up effects in the hippocampus and lateral temporal lobe (LTL). Left: for hippocampus, remembered trials showed significantly positive slopes (t(12) = 3.580, p = 0.004), whereas forgotten trials did not (t(12) = −0.669, p = 0.516). The slopes for remembered trials were significantly greater than those for forgotten trials (t(12) = 3.400, p = 0.005). Right: for the LTL, neither remembered trials (t(12) = −1.408, p = 0.185) nor forgotten trials (t(12) = −0.437, p = 0.670) of the visual short-term memory (VSTM) task showed significant slopes against zero. The slopes did not differ between VSTM remembered and forgotten trials (t(12) = −0.707, p = 0.493). (c) Control analyses by restricting hippocampal ripples (as detected in the main text) to those with a duration ≥ 25 ms, the results found a significant time ×VSTM accuracy interaction (β = 0.033, z = 2.983, p = 0.003). Further analyses revealed a significant ramping-up of ripple rates over the maintenance period for remembered trials (β = 0.022, z = 6.178, pFDR < 0.001), but not for the forgotten items (β = −0.010, z = −0.874, pFDR = 0.382), consistent with Fig. 3a in the main text. (d) Control analyses for hippocampal ramping-up effect using an alternative frequency range (80-120 Hz) and detection criteria (duration ≥ 25 ms and peak amplitude > 3 SD above baseline) following Vaz et al. (2019, Science). Our results found a significant time ×VSTM accuracy interaction was again observed (β = 0.024, z = 3.281, p = 0.001). Further analyses revealed a significant ramping-up of ripple rates over the maintenance period for remembered trials (β = 0.012, z = 4.998, pFDR < 0.001), but not for the forgotten items (β = −0.012, z = - 1.648, pFDR = 0.099), consistent with Fig. 3a in the main text. (e) Both fast and slow remembered trials showed significant hippocampal ramping-up effects (fast: β = 0.010, z = 2.463, pFDR = 0.014; slow: β = 0.018, z = 4.625, pFDR < 0.001), with no significant time × retrieval speed interaction (β = 0.009, z = 1.570, p = 0.117). The y-axis shows model-estimated ripple rates. (f) Both novel and repeated trials showed significant hippocampal ramping-up effects (novel: β = 0.018, z = 4.00, pFDR < 0.001; repeated: β = 0.009, z = 2.973, pFDR = 0.003), with no significant time ×novelty (novel vs. repeated) interaction effect (β = - 0.009, z = −1.716, p = 0.086). (g) Across all trials, channels, and participants, a linear mixed-effects model revealed a non-significant trend of decreasing LTL ripple rates during the maintenance period (β = −0.005, z = −1.772, p = 0.076). The y-axis shows model-estimated ripple rates. (h) LTL ripple rates decreased significantly for the fast trials, with a similar but non-significant trend for the slow trials (fast: β = −0.009, z = −2.245, pFDR = 0.049; slow: β = - 0.003, z = −0.721, pFDR = 0.471) and no significant time ×retrieval speed interaction (β = 0.006, z = 1.087, p = 0.277). The y-axis shows model-estimated ripple rates. (i) Control analyses when restricting analyses to LTL bipolar channel pairs with at least one contact located within LTL gray matter or within 2 mm of gray matter (n = 89 channels). The results revealed a significant time ×VSTM accuracy interaction was observed (β = −0.020, z = - 2.025, p = 0.043). Further analyses showed that this interaction was driven by a significant decrease in ripple rates for remembered items (β = −0.009, z = −2.641, pFDR = 0.017), whereas no significant effect was observed for forgotten trials (β = 0.012, z = 1.134, pFDR = 0.257). (j) The standard deviation of hippocampal ripple rates was significantly greater than that of LTL channels during the maintenance and pre-retrieval response periods across participants (all pscluster < 0.032), suggesting a greater heterogenous for hippocampal channels than LTL channels. Significant clusters are indicated by horizontal black bars. *: pFDR < 0.05, ns: not significant. *: p (pFDR) < 0.05, **: p (pFDR) < 0.01, ***: p (pFDR) < 0.001.
Control analyses for decoding accuracy in the lateral temporal lobe (LTL) for remembered and forgotten trials.
(a) Decoding accuracy was computed separately for each 1-s time bin of the maintenance period for remembered trials, with significant above chance (i.e., 25%) for the first, third, and fourth time bins (psFDR < 0.042). (b) Decoding accuracy of forgotten trials compared to chance level (25%) across the task (left: encoding and maintenance. Clusters showing significantly above-chance decoding accuracy are circled by black lines (all pscluster < 0.028). (c) Mean decoding accuracy across all time windows within each stage (encoding, maintenance, retrieval) was significantly above chance across participants (all psFDR < 0.037). (d) Decoding accuracy for VSTM remembered trials was significantly higher than for forgotten trials during retrieval in the cluster circled by black lines (pcluster = 0.012), with a similar trend but non-significant cluster during encoding (pcluster = 0.064, circled by grey lines). (e) Mean decoding accuracy across all time windows within each stage did not differ significantly between VSTM remembered and forgotten trials (all psFDR > 0.258). *: pFDR < 0.05, ns: not significant.
Control analyses for coupled ripple-locked hippocampal memory reactivation or independent ripple-locked memory reactivation
(a) HPC decoding accuracy compared to chance level (25%). Clusters indicate time windows with significantly above-chance decoding (pscluster < 0.031) and are circled by blacklines. (b) Averaged HPC decoding accuracies across all train-test time windows within individual task stages were all significantly above chance (psFDR < 0.038). (c) No significant clusters were found for HPC decoding accuracy locked to hippocampal-LTL coupled ripples across all task stages or within any task stage (all pscluster > 0.470). (d-e) Neither hippocampal (a) nor LTL (b) decoding accuracy was significantly locked to independent hippocampal ripples (all pscluster > 0.222). (f-g) Neither hippocampal (c) nor LTL (d) decoding accuracy was significantly locked to independent LTL ripples (all pscluster > 0.203). *: pFDR < 0.05.
Control analyses for ruling out long-term memory (LTM) confounds in hippocampal ramping-up effects, coupled ripples, and ripple-locked memory reactivation.
(a) Hippocampal ripples showed significant ramping-up effects for both subsequently LTM-remembered (β = 0.017, z = 4.141, pFDR < 0.001) and LTM-forgotten (β = 0.011, z = 2.987, pFDR = 0.003) trials, with no significant time ×LTM accuracy interaction (β = 0.003, z = 0.439, p = 0.661). (b) Hippocampal-LTL coupled ripple rates did not differ between subsequently LTM remembered and forgotten trials (pscluster > 0.626). (c) Coupled ripple-locked memory reactivation also showed no significant difference between subsequently LTM remembered and forgotten trials (all pscluster > 0.135). **: pFDR < 0.01, ***: pFDR < 0.001.
