Human hippocampal reactivation of amygdala encoding-related gamma patterns during aversive memory retrieval

Abstract

Emotional memories require coordinated activity of the amygdala and hippocampus. Human intracranial recordings have shown that formation of aversive memories involves an amygdala theta-hippocampal gamma phase code. Yet, the mechanisms engaged during translation of aversive experiences into memories and subsequent retrieval remain unclear. Directly recording from human amygdala and hippocampus, here we show that hippocampal gamma activity increases for correctly remembered aversive scenes. Crucially, patterns of amygdala high amplitude gamma activity at encoding are reactivated in the hippocampus, but not amygdala, during both aversive encoding and retrieval. Trial-specific hippocampal gamma patterns showing highest representational similarity with amygdala activity at encoding are reactivated in the hippocampus during aversive retrieval. This reactivation process occurs against a background of gamma activity that is otherwise decorrelated between encoding and retrieval. Thus, phasic hippocampal gamma responses track the retrieval of aversive memories, with activity patterns apparently entrained by the amygdala during encoding.

Human intracranial recordings reveal that during aversive memory retrieval, memory-specific gamma activity patterns, shaped by the amygdala during encoding, are reactivated in hippocampus.

Introduction

Episodic memory is typically better for emotional than for neutral events¹. One central question is whether emotional episodic memory reflects an augmentation of the same neurobiological mechanism(s) underlying neutral memory, or whether a qualitatively different process is engaged. In humans, lesions of the hippocampus impair both neutral and emotional episodic memory, whereas selective lesions of the amygdala leave neutral memory intact and reduce emotional memory performance to the same level as that for neutral stimuli². Functional MRI studies, although measuring a hemodynamic signal and lacking a temporal resolution with millisecond precision, also point to emotional memory enhancement involving co-participation of the amygdala and hippocampus^3,4.

Using direct intracranial recordings from humans with drug-resistant epilepsy, we recently described coordinated activity between amygdala and hippocampus during aversive, but not neutral, memory encoding⁵. However, successful encoding of both neutral and aversive stimuli was associated with a common manifestation: increased gamma activity in the hippocampus. The additional component to emotional memory encoding was that amygdala theta activity appeared to coordinate hippocampal gamma power to emotional stimuli. Furthermore, successful emotional encoding depended on the amygdala theta phase at which hippocampal gamma peaked; the phase difference, when considered in the time domain, was related to the lag between peaks of amygdala and hippocampal gamma activity. Thus, the contribution of the amygdala to successful emotional encoding includes a modulation of, and temporal coordination of, hippocampal gamma activity. This unique contribution suggests the presence of a qualitatively different process during emotional and neutral episodic memory.

Recently, the contents of memory have been probed using reinstatement analyses⁶, in which the pattern of activity evoked during retrieval is compared to that measured at encoding to test the similarity between these patterns both at the single stimulus and category (e.g., emotional vs neutral) levels. Converging evidence using functional MRI^7,8, electrophysiological^9,10 and magnetoencephalography measures^11–13 have identified encoding-retrieval similarity: neuronal activity patterns present during encoding that are reactivated during retrieval. Conversely, there is growing evidence for putative alterations in memory representations within various brain regions and memory stages^14–17. Thus, some questions remain regarding the interpretation of reactivation-related activity, such as whether BOLD signal measuring patterns of population activity could fail to isolate more transient^14–19 activity related to the content of memory. An additional question is whether reactivation patterns spread into different brain areas during a period of putative consolidation rather than remaining focal^17,20.

Here, we investigate amygdala and hippocampal oscillatory responses during the recognition of aversive scenes, using intracranial recording in patients from whom emotional memory encoding mechanisms were derived⁵. We further examine whether amygdala and hippocampal responses during recognition reflect reactivation²¹ of encoding patterns. Reinstatement analyses showed decorrelated encoding and retrieval activity in amygdala and hippocampus selectively for emotional stimuli. However, when representational patterns for remembered emotional pictures were derived from phasic peaks of amygdala gamma activity during encoding, we observed reactivation of these patterns solely in the hippocampus, both during encoding and retrieval. Critically, trial-specific patterns of hippocampal activity, showing highest representational similarity to that of the amygdala during successful emotional encoding, were subsequently reactivated in the hippocampus during successful emotional retrieval. Mechanistically, our results suggest that the amygdala drives hippocampal patterns during encoding. These hippocampal patterns are later reactivated in the hippocampus during retrieval of aversive information.

Results

Twenty-three participants with drug resistant epilepsy (Supplementary Table 1) undergoing intracranial recordings encoded and retrieved aversive and neutral scenes. The encoding data from thirteen participants were previously published⁵ and seven new participants have been included in the current analysis. In the encoding session, participants viewed 120 scenes (80 neutrals and 40 aversive) for 0.5 s and were asked if the image pertained to an indoor or outdoor scene. At recognition testing 24 h later, an equal number of old and new aversive and neutral scenes were shown and participants made remember (R), familiarity (K for “know”) and new (N) responses (Fig. 1a, b). Adopting the same approach as in our previous study⁵, memory performance was assessed by comparing the rates of eRHit (0.40 ± 0.04 mean ± SEM) and eKHit (0.23 ± 0.03) with nRHit (0.28 ± 0.03) and nKHit (0.21 ± 0.02) responses. This analysis revealed an interaction between emotion (aversive vs neutral) and memory (R vs. K), (within-subjects two-way ANOVA, F_(1,22)  =  4.39, P  =  0.048, Fig. 1c, e). During both encoding and recognition sessions, we simultaneously recorded from the amygdala (n = 20 patients) and ipsilateral hippocampus (n = 14 patients) (Methods, Fig. 1d, Supplementary Fig. 1). Patients with fewer than four trials in any condition were excluded, resulting in n = 17 patients with amygdala electrodes and n = 12 patients with electrodes in both structures (Supplementary Table 2, Supplementary Table 3).

Fig. 1. Task design, experimental conditions including trials number and electrode contact localization.

a Task design with examples of emotional aversive (e) and neutral (n) scenes presented during the encoding and recognition session. b Schematic of the experimental conditions at recognition. Aversive and neutral scenes received “remember”, “know”, and “new” (R, K, N) responses. Old scenes can be classified as correctly remembered (eRHit, nRHit), correctly familiar (eKHit, nKHit) or miss (eMiss, nMiss). New scenes can be classified as remember false alarms (eRFA, nRFA), familiar false alarms (eKFA, nKFA) or correctly rejected (eCRj, nCRj). c Recognition performance (n = 23 participants) for emotional aversive (e) and neutral (n) correctly remember (R), remembered false alarms (RFA), correctly known (K) and known false alarms (KFA). Here, and in subsequent plots, each dot represents one patient’s data, and horizontal and vertical line reflects mean and standard error of the mean, respectively. Performance is expressed as % of number of trials, i.e., a score of 100% for the eR condition would mean that there were no false alarms and that every old aversive stimulus was correctly remembered. d Schematic representation of electrode contact localization illustrates the positions of contacts for the 20 participants included in the electrophysiological analysis in the left and right amygdala (A), represented in pink, along with contacts for the 14 participants in the ipsilateral hippocampus (H) represented in light blue. That is, patients with contacts in amygdala and ipsilateral hippocampus are labeled (A-H), or if contacts are limited to amygdala, they are labeled (A). e Memory performance comparing the rates of R and K responses for emotional versus neutral items, n = 23. Data are presented as mean values ±  SEM.

The hippocampus tracks the retrieval of aversive memories

A means to probe the contents of memory is to examine responses to previously encoded stimuli in the context of tests of recognition. Typically, responses are compared as a function of performance, including comparisons of correct hits vs. misses (items previously presented that receive “old” vs “new” responses, respectively) or hits vs. correct rejections (“new” responses of previously unpresented foils). These effects can then be compared between emotional and neutral trials (Fig. 1b).

We investigated retrieval-related induced responses in the hippocampus and the amygdala by comparing aversive and neutral scenes successfully recollected (eRHit and nRHit, respectively) with scenes that were not. Correct “know” (KHit) and incorrect “new” (Miss) responses were collapsed together (eKHit&eMiss, nKHit&nMiss) as similar power spectral results were observed comparing Hit remembered trials with either KHit or Miss trials (Supplementary Fig. 3a–c). Aversive scenes correctly remembered (eRHit) elicited higher gamma activity (60–85 Hz) in the hippocampus beginning 0.7 s after stimulus presentation and lasting until 1.1 s (emotion by memory interaction; summed t-value = 743.09, P = 0.014, n = 12, Fig. 2a–c). Post-hoc t-tests on mean power changes across the significant time-frequency cluster showed a difference for aversive (eRHit vs. eKHit&eMiss t₁₁ = 4.54, P  =  0.0001, d  =  1.31) but not neutral scenes (nRHit vs. nKHit&nMiss t₁₁ = 0.07, P  =  0.94, d  =  0.02, Fig. 2c). Similar hippocampal gamma activity (from 0.8 until 1.1 s, 63–90 Hz) was found when comparing the gamma spectral responses to aversive vs. neutral correctly remembered scenes with aversive vs. neutral KHit (emotion by memory interaction; summed t-value = 621.19, P = 0.035, n = 12, Supplementary Fig. 3b) or Misses (emotion by memory interaction; summed t-value = 475.04, P = 0.043, n = 12, Supplementary Fig. 3c, from 0.7 until 0.9 s, 60–83 Hz). Note that these latter two t-statistics are only significant for one-tailed comparison (two-sided threshold p < 0.025).

Fig. 2. The hippocampus tracks correctly remembered aversive scenes, whereas the amygdala responds to aversive vs. neutral scenes independently of memory performance.

Time-frequency plots of hippocampus (a–c) and amygdala (d–f) gamma responses (35–150 Hz) for aversive and neutral scenes correctly remembered (eRHit, nRHit) and correct “know” and incorrect “new” responses (eKHit&eMiss, nKHit&nMiss). b Time-frequency resolved test statistics for the comparison of the emotion by memory effect, two-sided paired t-test (cluster-based permutation test). The significant cluster is indicated by a black outline. c Scatter plot show the mean hippocampal gamma values in the significant cluster, relative to baseline, for each patient, n = 12. Horizontal and vertical line reflects mean and standard error of the mean. d same as for (c), but for the amygdala time-frequency data. e The statistical test shows the comparison aversive vs neutral and the significant cluster (black outline) after applying cluster-based permutation test. f Scatter plot shows the mean amygdala gamma power change (relative to baseline) in the significant cluster, n = 17. Horizontal and vertical line reflects mean and standard error of the mean.

At recognition, memory responses can also be assessed by comparing hit and correct rejections. There was a trend towards an interaction when hippocampal gamma activity for aversive vs. neutral correctly remembered scenes was contrasted with those that were correctly rejected (emotion by memory interaction; summed t-value = 441.59, P = 0.069, Supplementary Fig. 3a, d).

By contrast to the hippocampus, amygdala gamma activity was not significantly modulated by the retrieval of aversive scenes (i.e., an emotion by memory interaction was not observed). A significant increase in gamma power was found while participants observed aversive, but not neutral scenes (Fig. 2d–f). This broadband gamma effect (35–130 Hz) started at stimulus onset and lasted until 0.7 s. (main effect of emotion, cluster 1: summed t-value = 1084.9, P = 0.022, cluster 2: summed t-value = 903.03, P = 0.031; n = 17, Fig. 2e). Within the lower oscillatory frequencies range (1–34 Hz), no differences in condition were observed in the hippocampus (Supplementary Fig. 4a), whereas the amygdala showed a significant emotion by memory interaction in the theta range (4–6 Hz, from 0.4 until 1.3 s) (summed t-value = 493.11, P = 0.016, n = 17). However post-hoc t-tests showed that this was driven by a significant difference only for neutral scenes (Supplementary Fig. 4b).

Encoding-retrieval global gamma decorrelation selective for emotional memories

There is evidence that memory retrieval involves the reactivation of encoding-related neural activity^{7,8,10–13,22,23}. However, an alternative perspective posits that episodic memory operates as a constructive process, entailing neuronal reorganization over time and the transformation of memory content, manifesting as changes in representational patterns between different memory stages^14–17. Yet, it remains unclear whether amygdala and hippocampal responses observed during recognition reflect reactivation of encoding patterns.

To investigate whether successful emotional retrieval promoted reactivation of encoding activity patterns in amygdala and hippocampus, we employed representational similarity analysis (RSA)^10,24–26 between the same items presented at encoding and retrieval. Again, we sorted trials into four categories depending on whether participants remembered (R) the aversive or neutral scenes or whether the item received a correct “know”, or incorrect “new” response (KHit&Miss). Specifically, we computed the similarity between encoding and retrieval activity patterns (ERS; Methods, Fig. 3a) in the gamma range (between 35 and 150 Hz) in windows of 0.5 s, sliding in 0.4 s (80% overlap). We used Spearman’s rho correlation as our similarity metric and computed the ERS in the amygdala and the hippocampus separately.

Fig. 3. Decorrelated encoding-retrieval gamma (35–150 Hz) activity for aversive scenes in the hippocampus and amygdala.

a Schematic depiction of the encoding - retrieval similarity (ERS) analysis. Two example epochs of 0.5 for encoding and retrieval are shown. The representational pattern is presented for both sessions as a one-dimension vector which concatenates the frequency (35–150 Hz in 2.5 Hz steps) and time dimensions. For each trial, the Spearman correlation was computed between encoding and retrieval from −0.5 to 1.5 s in time windows of 0.5 s, overlapping by 0.4 s, and the correlations averaged within-condition. The encoding-retrieval reactivation map is shown in an example trial, where values in the diagonal represent the reactivation without lag (i.e., the same time period at encoding and retrieval), while off-diagonal values depict lagged correlations. b Left: Hippocampus grand average encoding-retrieval reactivation map for aversive (e) and neutral (n) successful encoded and correctly remembered (RHit) for items that received a known/new response (KHit&Miss). Right top: results of the test for the interaction (eRHit-eKHit&Miss) – (nRHit-nKHit&Miss) are shown with the significant cluster outlined in black, n = 12. The color bar pertains to the correlation coefficient. Right bottom: the mean ERS values for each patient and the four trial types are plotted. Horizontal and vertical line reflects mean and standard error of the mean, respectively. c Same as (b) but for the amygdala ERS analysis, n = 17. Horizontal and vertical line reflects mean and standard error of the mean.

Strikingly, this analysis revealed a negative correlation between gamma activity patterns during encoding and retrieval only for aversive successful encoded and later remembered scenes in both the hippocampus (emotion by memory interaction, summed t-value = −55.73, P = 0.048, n = 12, Fig. 3b, eRHit mean ± SEM -0.027 ± 0.017; eKHit&Miss 0.020 ± 0.011; nRHit 0.004 ± 0.009; nKHit&Miss -0.003 ± 0.008) and the amygdala (emotion by memory interaction, summed t-value = -88.35, P = 0.017, n = 17, Fig. 3c, eRHit mean ± SEM −0.037 ± 0.014; eKHit&Miss 0.014 ± 0.009; nRHit 0.018 ± 0.010; nKHit&Miss 0.0002 ± 0.068). As a control analysis, we investigated potential frequency band-dependent reactivation effects by taking a subset of frequency ranges within the 35–150 Hz range, and repeating the analysis restricting to high gamma (90–150 Hz) and low gamma (60–85 Hz) ranges (Supplementary Fig. 5). The 60–85 Hz range includes the observed higher gamma activity in the hippocampus for correctly remembered aversive scenes, whereas the 90–150 Hz range was used to determine amygdala gamma peaks at encoding and define representational patterns reactivated at retrieval. None of these analyses revealed a positive correlation in either structure. In the amygdala, high gamma range (90–150 Hz) patterns showed negative correlation between encoding and retrieval, specifically for aversive scenes that were successfully encoded and later remembered, mirroring effects observed when taking a broad gamma range (35–150 Hz), (Supplementary Fig. 5b). However, this effect did not reach corrected significance, nor did effects in hippocampus (Supplementary Fig. 5a), raising a possibility of insufficient statistical power in both structures when the full gamma range is not included. We did not observe any significant effect in the low-frequency range (1–34 Hz) in either structure (Supplementary Fig. 5a, b) supporting selectivity of reinstatement to the gamma range.

Encoding-retrieval global gamma decorrelation is not observed in lateral temporal cortex

The gamma pattern decorrelation during the encoding and retrieval of aversive, later-remembered scenes was observed in both regions tested (the amygdala and hippocampus). To provide a measure of specificity of this decorrelation to these structures, we repeated the same analysis using channels in the lateral temporal cortex on the same electrode reaching the medial temporal lobe (Supplementary Fig. 2). We did not observe a significant decorrelation in lateral temporal cortex (Supplementary Fig. 6). To confirm the specificity of the negative correlation effects in the amygdala and hippocampus compared to the lateral temporal cortex, we extracted the mean encoding-retrieval similarity (ERS) values for each of the four conditions (eR, eKF, nR, nKF) from the significant clusters identified in the amygdala and hippocampus. The same procedure was applied to the lateral temporal cortex, using the overlapping time and frequency data points from the amygdala and hippocampus effects. We then computed the interaction effect within each region and performed a paired t-test to directly compare the effects across the three brain regions. The negative correlation observed in both the amygdala and hippocampus was significantly different from that in the lateral temporal cortex (amygdala mean -0.067 ± 0.026 vs. lateral temporal cortex mean 0.014 ± 0.016: t₁₁ = −2.44, P = 0.032; hippocampus mean −0.056 ± 0.023 vs. lateral temporal cortex: t₁₁ = -3.18, P = 0.008). In contrast, the negative correlation in the amygdala did not significantly differ from that in the hippocampus (t₁₁ = −0.33, P = 0.74) (Supplementary Fig. 7). These findings suggest that, at least at a macro-level of observation that is induced responses of the local field potential, gamma activity induced at encoding is decorrelated from that induced by the same emotional stimulus at retrieval in both amygdala and hippocampus, but not in the lateral temporal cortex.

Encoding-related amygdala patterns around gamma peaks are reactivated only in the hippocampus at encoding and retrieval

Evidence from human recordings suggests that cell assemblies firing at particular frequencies are responsible for encoding episodic events and reinstated during retrieval^27,28. Gamma bursts have been previously linked to neuronal spiking in monkeys during a working memory task²⁹. During aversive memory formation we observed a time-lagged correlation between amygdala and hippocampal gamma peaks⁵. We thus reasoned that the reactivation process for aversive memory, which was not evident at the level of induced gamma averaged over encoding-retrieval trial pairs, could be better assessed by analyzing neuronal patterns aligned to gamma peaks in the amygdala during encoding.

The modulation of memory by emotion is generally thought to be a time-dependent process, requiring a period of consolidation^30,31, although some studies show emotion-induced memory enhancement after only brief delays³². The latter observation therefore raises a question as to whether amygdala encoding activity patterns might already be reflected in hippocampal activity patterns during the encoding phase.

To address both questions, we employed representational similarity analysis to probe (1) encoding-retrieval similarity (ERS) between patterns of activity in the amygdala at encoding and amygdala (Supplementary Fig. 8) or hippocampus at retrieval (Fig. 4e); (2) encoding-encoding similarity (EES) analysis between patterns of activity in the amygdala at encoding and the hippocampus at encoding (Fig. 4f). (3) As a final control analysis, we tested for ERS between patterns of activity in hippocampal gamma peak at encoding and hippocampal activity during retrieval (Supplementary Fig. 10).

Fig. 4. Representational patterns around amygdala gamma peaks at encoding are reactivated in the hippocampus during encoding and retrieval of recollected aversive scenes.

Schematic representation of the analysis (a–d). a Left: Example of one trial showing broadband gamma, bandpass filtered between 90–150 Hz (time pertains to picture onset). Red downwards arrows point to gamma peaks detected in the amygdala (see Methods). We extracted a time-resolved spectral activity pattern for each amygdala peak by selecting the activity 0.05 s before and after the peak. The semi-opaque vertical rectangle illustrates this selection for one specific peak. Right: time-resolved representational pattern at second peak of the example trial. b Left: Hippocampal time-frequency induced response to the same picture during retrieval, along with a single time-resolved spectral activity representational pattern corresponding to the first 0.1 s (semi-opaque vertical rectangle) after picture onset, right. c Left: Hippocampal time-frequency induced response to the same picture during encoding (i.e., the same trial as in a), along with a single representational pattern locked to the time of the amygdala peak, right. The semi-opaque vertical rectangle represents the hippocampal data windowed 0.05 s before and after the time of the amygdala peak shown in (a). d Each representational pattern was further converted to a one-dimension vector containing frequency (90–150 Hz in 2.5 Hz steps) and time (0.1 s) resolution. For each trial, in each condition, Spearman’s correlation was computed between one pattern in the amygdala and another pattern in the region of interest (amygdala or hippocampus) over time. e Encoding-retrieval similarity analysis (ERS) results: the correlation coefficient for eRHit and eKHit&Miss conditions (left panel) and for nRHit and nKHit&Miss (middle panel) are plotted. Data are presented as mean values ± SEM. The x-axis represents the time in the hippocampus from −0.5 to 1.5 s. T-values are plotted over time. The significant cluster (cluster-based corrected) for the emotion by memory interaction is shown in red (right panel). f Encoding-Encoding similarity (EES) analysis between amygdala and hippocampus is displayed as in (e). Here, the x-axis represents time relative to amygdala peaks (0) and from −0.5 until 1.5 s after the peak.

To evaluate these hypotheses, we first identified peaks of high amygdala gamma activity (90–150 Hz) during encoding in each trial for each condition from 0.4 to 1.1 s (Methods, Fig. 4a). This is the window in which we observed a significant gamma effect for successful encoding of aversive scenes⁵. Hippocampal reactivation of amygdala activity patterns observed during gamma bursts may be triggered by different events during encoding and retrieval. We hypothesized that during encoding, gamma bursts directly drive hippocampal activity with a short latency. Therefore, we identified gamma peaks in the amygdala and computed EES within a time window from 0.5 s before to 1.5 s after each peak (Fig. 4c). In contrast, during retrieval, amygdala activity patterns associated with gamma bursts at encoding may be reactivated after a delay following stimulus presentation. This delay may not correspond to the latency of gamma bursts relative to stimulus onset during encoding. Thus, we adopted a different approach for the ERS analysis, examining activity in a time window from 0.5 s before to 1.5 s after stimulus onset (Fig. 4b).

Notably, amygdala representational patterns at encoding were not reactivated in the amygdala at retrieval (Supplementary Fig. 8). By contrast, the ERS analysis revealed a time-period post-stimulus onset from 0.5 to 0.7 s and from 0.9 to 1.1 s in which amygdala encoding activity patterns were reactivated in the hippocampus only for the aversive scenes correctly recognized (emotion by memory interaction, summed t-value = 39.19, P = 0.023; second cluster, summed t-value = 38.68, P = 0.024, n = 12, Fig. 4e). Single subjects’ results for the amygdala-hippocampus ERS and EES analysis are reported in Supplementary Figs. 9, 11, respectively. We did not observe the same emotion by memory interaction when hippocampal activity during retrieval was locked to the hippocampal gamma peak at encoding (Supplementary Fig. 10). This lack of effect supports the proposed mechanism, emphasizing the crucial role of the amygdala, rather than the hippocampus, in driving the reinstatement process. This process involves representational patterns encoded by the amygdala that are reactivated in the hippocampus during subsequent retrieval of the same information.

The EES analysis showed patterns of hippocampal gamma activity that mirrored that of amygdala gamma peaks occurring 0.5 s to 1 s earlier, specifically for aversive scenes later remembered (emotion by memory interaction, summed t-value = 65.62, P = 0.0068, n = 12, Fig. 4f). This result is in line with growing evidence, mainly from human fMRI and scalp EEG studies, that post-encoding awake reactivation strengthens memory³³.

Critically, the reactivation of encoding patterns in the amygdala and the hippocampus at encoding (EES) and retrieval (ERS) was specific to the representational patterns locked to the amygdala peaks, since no significant effects were observed when the signal was locked to random ‘non-peak’ time periods within each individual trial. Results were consistent over 1000 repetitions of random non-peak selection (P-values for the 1000 random selections are reported for the EES analysis in Supplementary Fig. 12a and ERS in Supplementary Fig. 12b).

To further validate the robustness of the observed ERS and EES effects and rule out potential confounds related to the selection of the peaks, we conducted additional control analyses. First, we repeated these tests but increasing the minimum inter peak distance from 0.1 to 0.3 s to account for the autocorrelation of the gamma band activity³⁴ in the amygdala and replicated our original results (ERS in the hippocampus: emotion by memory interaction, summed t-value = 61.18, P = 0.0058; second cluster, summed t-value = 54.44, P = 0.0092, n = 12, Supplementary Fig 13a, EES in the hippocampus: emotion by memory interaction, summed t-value = 52.65, P = 0.012, n = 12; Supplementary Fig. 14a). Secondly, instead of considering all detected amygdala gamma activity peaks in a trial, we randomly selected one peak per trial. The ERS analysis between amygdala and hippocampus was close to significance (summed t-value = 18.38, P = 0.075, n = 12; Supplementary Fig. 13b), potentially suggesting that the information carried by more than one peak may be needed for the reactivation process. The EES analysis showed a significant emotion by memory interaction (summed t-value = 74.27, P = 0.0024; n = 12, Supplementary Fig. 14b). Finally, we repeated the EES analysis and tested whether early and late amygdala peaks showed similar EES with hippocampal pattern and found a similar trend in both results (Supplementary Fig. 15). When including only late peaks, the emotion by memory interaction (summed t-value = 29.38, P = 0.039, n = 12) survived one-tailed correction (two-sided t-test threshold p < 0.025).

Hippocampus patterns at encoding are reactivated during retrieval of aversive scenes

Amygdala high amplitude gamma patterns during encoding reactivated in the hippocampus at encoding and retrieval. A question remained as to whether the hippocampal patterns that mirrored amygdala activity during encoding were also reactivated during the retrieval process. Consequently, we examined reactivation of hippocampal patterns reactivation between encoding and retrieval as a function of the amygdala- hippocampus representation pattern similarity found at encoding. To do so, we used the correlation coefficient score obtained in the encoding-encoding similarity analysis (EES) between amygdala and hippocampus over time. First, for each amygdala pattern derived around amygdala gamma peaks at encoding, we obtained a correlation coefficient reflecting the similarity of this pattern with the hippocampal pattern at encoding over time. Secondly, we determined the time window of maximum similarity between amygdala and hippocampal representational pattern by selecting the time at which the highest correlation coefficient occurred. We next calculated the delay between amygdala gamma peak at encoding and this time point of highest correlation between amygdala and hippocampus and centered hippocampal representation patterns in time to this event (+/-0.05 s). Finally, we employed the representational similarity analysis between hippocampal patterns at encoding and hippocampal patterns at retrieval detected from stimulus onset until 1.5 s, Fig. 5a–d. A significant emotion by memory interaction was observed (summed t-value = 36.16, P = 0.021, n = 12, Fig. 5e), with the largest difference between conditions identified between 0.7 and 0.9 s after stimulus onset. This finding demonstrates that the hippocampal gamma patterns from the encoding phase that are subsequently reactivated during the retrieval of aversive scenes are the patterns present when there is maximal pattern similarity between the amygdala and hippocampus during encoding.

Fig. 5. Patterns of hippocampal gamma activity at encoding that are most similar to amygdala emotional encoding activity are reactivated in the hippocampus during emotional retrieval.

Schematic representation of the analysis (a–d). a Example of one subject correlation coefficient of the encoding-encoding similarity (EES) between amygdala and hippocampus over time relative to amygdala gamma peak (y axis). Each row is a peak detected in the amygdala at encoding (y axis). b Example of correlation coefficients between amygdala and hippocampus at encoding derived from one amygdala gamma peak. The red triangle points to the highest correlation coefficient. c Hippocampal time-frequency data for the same trial where the amygdala peak was detected at encoding. The red vertical line represents the time where the maximum similarity between representational patterns in the two brain structures during encoding was detected (time of the peak + the delay of maximum similarity between the two structures), along with an example of time-frequency data windowed (gray shadow) 0.05 s before and after this event. d Example of representational patterns included in the RSA analyses. Patterns were converted into a one-dimension vector and the Spearman correlation was computed between one pattern in the hippocampus at encoding and another pattern in the hippocampus at retrieval over time (0.8%-time overlap). e Encoding-retrieval similarity analysis (ERS) n = 12: data show the correlation coefficient for eRHit and eKHit&Miss conditions (left panel) and for nRHit and nKHit&Miss (middle panel) from −0.5 to 1.5 s. Data are presented as mean values ± SEM. In the right panel, T-values are plotted over time. The x-axis represents the time in the hippocampus at retrieval from stimulus onset (0) to 1.5 s where the statistical test was performed. The significant cluster (cluster-based corrected) for the emotion by memory interaction is shown in red.

Discussion

Events with emotional significance are generally more memorable compared to neutral ones³⁵. However, it is unclear whether fundamentally distinct neurobiological mechanisms are engaged for neutral vs emotional memory formation and retrieval. Both the amygdala and hippocampus are involved in encoding and retrieval of emotional memories^4,5,36, but currently, only direct and simultaneous recording from both structures can reveal the mechanisms underlying their dynamic interplay in humans. Through the use of simultaneous intracranial recordings from the human amygdala and hippocampus, we found that during retrieval of previously seen images, hippocampal gamma activity was higher for aversive scenes correctly remembered, whereas the amygdala only showed a general response to aversive vs neutral scenes.

The question of whether memory retrieval involves a reactivation process^8,11 or rather a transformation or reconstruction of the memory trace¹⁴ is a subject of ongoing debate¹⁷. The perspective influenced by the concept of ‘mental time travel'³⁷, emphasizes the reactivation of encoded memory representations^10,12,22,23. By contrast, an increasing number of studies have reported putative transformations in memory representations when examining representational contents in different brain regions between encoding and retrieval^14–17. Alternatively, one may argue that different approaches may isolate different aspects of the reactivation process. Neuronal activity could be decorrelated at the macro-level, but the memory content could be better isolated by analyzing neuronal patterns aligned with more transient neuronal activity.

Our first analysis of encoding-retrieval similarity within the gamma frequency range (35–150 Hz) showed that for correctly remembered aversive scenes, both the amygdala and the hippocampus exhibited a decorrelation between encoding and retrieval representational patterns when compared to the rest of trial types (Fig. 3). This finding may reflect the mismatch between gamma frequencies that are employed during successful encoding⁵ (emotion by memory interaction in the amygdala was observed at around 120 Hz) and successful retrieval of aversive scenes (emotion by memory interaction in the hippocampus was observed at around 70 Hz). Thus, this approach did not appear to isolate a reactivation process, instead indicating that– at a global level – gamma activity is different under emotional encoding vs retrieval contexts. An increasing number of studies have reported putative transformations in memory representations when examining representational contents in different brain regions between encoding and retrieval^14,15,17. Nevertheless, observing decorrelation is not sufficient to infer a transformation process.

Strikingly, when we restricted the same analysis approach to periods of high amplitude gamma activity, we discovered that these phasic neuronal patterns in the amygdala at encoding did indeed reactivate but only in the hippocampus and only for aversive remembered scenes. These patterns of neuronal activity originating in the amygdala during encoding were reactivated in the hippocampus during both the encoding, around 0.4 s after stimulus offset, and retrieval of the same information. The latency of encoding reactivation is consistent with findings from scalp EEG studies, showing a higher degree of reactivation of encoding patterns to pictures later successfully vs unsuccessfully recalled during the post-encoding period, around 0.5 s from stimulus offset^19,38.

Importantly, the phasic hippocampal gamma patterns that were detected as reactivations of amygdala gamma peaks during encoding, were also reactivated in the hippocampus at retrieval. These results suggest that phasic patterns of amygdala neuronal activity to successfully encoded emotional pictures are mirrored by hippocampal gamma activity (after some lag) during encoding and these encoding-related patterns are reproduced 24 h later during the recall of aversive memory^27,39,40. That is, the reactivation process of phasic amygdala gamma patterns is initiated in the hippocampus during the encoding phase and results in the reactivation of hippocampal patterns during retrieval in the hippocampus, but not amygdala.

Representational similarity analysis approaches applied to fMRI data have shown representational reinstatement during the retrieval of emotional memories in the human brain⁴¹. The application of RSA to iEEG data in our study critically complements this previous work by demonstrating the crucial role of amygdala and hippocampal gamma frequency activity in the reactivation of emotional memories. One key advantage of applying RSA to iEEG data, compared to fMRI, is the fine-grained temporal insight it offers into the specific spectral frequencies involved in memory reactivation, which might vary depending on the particular representational content and memory functions studied. Indeed, previous iEEG investigations revealed the crucial role of gamma activity in the encoding of item-specific representations⁴², beta oscillations in the representation of items during visual working memory in the PFC⁴³ and theta oscillations in the encoding of hippocampal item-context associations¹⁰. Our study extends this body of work by showing the reactivation of representational signals locked to amygdala gamma peaks, a reinstatement process that cannot be selectively captured by current fMRI approaches.

Understanding how re-exposure to aversive stimuli reactivates encoding patterns after 24 h is of potential clinical relevance, for example for post-traumatic stress disorder (PTSD), characterized by emotion dysregulation and fear memory^44,45, or anxiety disorders. The mechanisms underlying aversive memory reactivation is also relevant to potential therapeutic avenues aimed at rendering these memory traces susceptible to pharmacological⁴⁶ or behavioral interventions. Furthermore, providing a better mechanistic understanding of amygdala-hippocampal communication during encoding, reactivation and retrieval of aversive information – with precise temporal and spectral resolution – can inform the design of novel stimulation protocols that specifically reproduce or disrupt their dynamic communication. These protocols can build on previous evidence that amygdala theta-burst stimulation enhances memory for neutral stimuli⁴⁷, and a recent report that amygdala closed loop neuromodulation reduces symptoms in two PTSD patients⁴⁸.

In summary, our findings provide evidence of a progression of the aversive memory trace over time and shed light on the role of amygdala-hippocampal dynamics that facilitate successful storage and retrieval of aversive memories. This dynamic process is evident in the dissociable responses of the amygdala and hippocampus during memory encoding and retrieval, and opposite directionality of mutual influence during the latter two processes. Critically, our findings demonstrate that amygdala patterns at encoding, detected around gamma peaks, are reactivated in the hippocampus during both the successful encoding of aversive information and its subsequent retrieval. That is, despite a lack of global correlation in gamma activity between the encoding and retrieval of aversive scenes in both the amygdala and the hippocampus, the peak gamma activity patterns originating in the amygdala during encoding are reactivated in the hippocampus during the encoding and retrieval phase. Moreover, hippocampal encoding patterns that mirrored amygdala activity during encoding, at the time of maximum similarity between gamma activity in the two structures, were also reactivated in the hippocampus during emotional memory retrieval process. The overall mechanism emphasizes the prominent role of the amygdala in aversive memory formation. This suggests that the amygdala drives hippocampal activity during encoding, which, in turn, results in representational patterns that are reactivated in the hippocampus during subsequent retrieval of the same information. These observations have the potential to guide the development of more effective treatments aimed at targeting the amygdala-hippocampal network, ultimately with the goal of enhancing or modifying the content of memory traces.

Methods

Participants

In the current dataset 23 patients (Supplementary Table 1) with refractory focal epilepsy (12 female) participated, with depth electrodes implanted only for diagnostic purposes. All participants were right-handed, except for one who was left-handed. All participants had electrodes in the amygdala and 14 also in the ipsilateral hippocampus. 16 participants were recorded at the Hospital Ruber International, Madrid and 7 at the Swiss Epilepsy Center, Zurich, Switzerland. All participants signed informed consent and did not receive financial compensation. The study received full approval from the local ethics committees of the Hospital Ruber Internacional, Madrid, Spain and Kantonale Ethikkommission, Zurich, Switzerland (PB-2016-02055). Exclusion criteria: Three participants were subjected to exclusion from the electrophysiological analysis on the grounds of insufficient signal quality. Additionally, two participants with electrodes in both the amygdala and hippocampus were excluded because they had fewer than four trials in at least one condition. Similarly, one further participant with the electrode only in the amygdala was excluded based on the same criterion (Supplementary Table 2, Supplementary Table 3). Results of encoding related responses of 13 of the 23 participants presented here have been previously reported⁵.

Stereotactic electrode implantation

We conducted a pre-operative contrast-enhanced MRI under stereotactic conditions to map vascular structures before electrode implantation for participants recorded in Madrid at the Hospital Ruber Internacional. This process allowed to calculate stereotactic coordinates for electrode trajectories using the Neuroplan system (Integra Radionics). We used DIXI Medical Microdeep depth electrodes, which are multi-contact, semi-rigid with a diameter of 0.8 mm, contact length of 2 mm, and an inter-contact isolator length of 1.5 mm. For the majority of patients, the implantation followed the stereotactic Leksell method, while three patients underwent implantation using the Medtronic Robot-assisted procedure. For participants recorded in Zurich, we used Ad-Tech depth electrodes with a diameter of 1.3 mm, contact length 1.6 mm, and a spacing of 3 mm between the most medial contact centers. Robot-assisted stereotactic implantation was conducted based on direct targeting in contrast enhanced MRI. Lead localization was verified in post-operative MRI. Depth electrodes were stereotactically implanted into the amygdala, hippocampus, and entorhinal cortex.

Data acquisition

In Madrid, the intracranial EEG (iEEG) activity was obtained by utilizing an XLTEK EMU128FS amplifier manufactured by XLTEK, located in Oakville, Ontario, Canada. The iEEG data were recorded at a rate of 500 Hz at each electrode contact site and they were referenced to linked mastoid electrodes. Intracranial data at the Swiss Epilepsy Center were recorded using the Neuralynx ATLAS system with a sampling rate of 4096 Hz and an online band-pass filter range of 0.5 Hz to 1000 Hz, all against a common intracranial reference. All data with sampling rate above 500 Hz were down-sampled to 500 Hz.

Electrode contact localization

To locate the electrodes, we co-registered the post-electrode placement CT scans (post-CT) with the pre-electrode placement T1-weighted magnetic resonance images (pre-MRI) for each patient. To optimize co-registration, we initially skull-stripped both brain images. For CTs, we filtered out voxels with signal intensities between 100 and 1300 HU. For pre-MRI skull stripping, we first normalized the image to MNI space using the New Segment algorithm in SPM8. The resulting inverse normalization parameters were then applied to transform the brain mask into the native space of the pre-MRI. Voxels outside the brain mask with signal values in the highest 15th percentile were filtered out. We then co-registered and re-sliced the skull-stripped pre-MRI to the skull-stripped post-CT. Subsequently, the pre-MRI was affine normalized to the post-CT, effectively transforming the pre-MRI image into native post-CT space. The two images were overlaid, and the post-CT was threshold to visualize only the electrode contacts.

Electrode contact visualization

To visualize the electrodes using Lead-DBS⁴⁹ (lead-dbs.org), we initiated the process by selecting the electrode model for each patient (Dixi or Ad-Tech). We then performed co-registration of CT scans with MRI using the advanced normalization tools (ANTs), and subsequently, the volumes were normalized into MNI ICBM 2009b nonlinear asymmetrical space based on the preoperative MRI data. Additionally, the software was used to correct for any brain shift that occurred. We initially pre-reconstructed the electrodes using manual reconstruction. The reconstructed electrodes were visually inspected, and if any misalignments were detected, we manually adjusted the reconstruction. This adjustment was based on postoperative CT, ensuring the trajectory was placed as accurately as possible within the center of the electrode artifact. To visualize all the electrodes, we employed the Lead-group software⁵⁰. After this procedure, MNI coordinates were assigned to each electrode contact. In the final step, we plotted only the contacts for each electrode within the regions of interest, either the amygdala or the hippocampus.

Stimuli

During the encoding session, participants were exposed to a total of 120 pictures—40 with aversive content and 80 with neutral content. These pictures were selected randomly from a larger pool of stimuli, which included 80 high-arousing aversive scenes sourced from the International Affective Picture System (IAPS)⁵¹ and 160 low-arousing neutral pictures. Aversive scenes involve mutilations, attacks and blood. Of the neutral images, 149 were chosen from the IAPS (depicting household scenes and neutral individuals), while the remaining eleven neutral landscape pictures were sourced from the internet. On a nine-point scale measuring valence, the mean normative IAPS ratings (standard error of the mean) were 5.05 (±0.05) for neutral pictures and 2.04 (±0.05) for aversive pictures. For arousal, the corresponding ratings were 3.29 (±0.06) for neutral pictures and 6.3 (±0.07) for aversive pictures. Given that emotional items tend to be more memorable than neutral ones, the ratio of aversive to neutral stimuli was deliberately set at 1:2 to ensure a balanced number of trials per condition.

Task

Participants encoded and retrieved after 24 h aversive and neutral scenes selected for the International Affective Picture System (IAPS) dataset⁵¹.

The encoding and recognition sessions took place on the third and fourth post-operative days, respectively, in Madrid (Spain), and on the second and third post-operative days in Zurich (Switzerland). Before signing the informed consent, patients were shown an example of an aversive IAPS picture and informed that they would see similar images on that day and the next. Task instructions were provided verbally and on-screen, either in Spanish or in German for patients recorded in Switzerland. During the encoding session, both emotional and neutral pictures were presented in a pseudo-random order (display time: 0.5 s; interstimulus interval: 3.5 s) with the constraint that each emotional picture was followed by at least one neutral picture. The pictures were displayed on a 27 × 20.3 cm video monitor (resolution: 1024 × 768 pixels) positioned approximately 50 cm from the patient’s eyes (visual angle: 30.2° × 22.9°). Patients were instructed to make an indoor-outdoor judgment for each picture using a button press.

On the second day, after 24 h, they underwent a memory test in which they classified 240 scenes (80 neutral and 40 aversive old, 80 neutral and 40 aversive new) as remembered (R), familiar/know (K) or new (N). During encoding participants were not informed about the memory test and they were just asked to perform an indoor-outdoor judgment. During both sessions, patients were asked to remain as still as possible, focus on the center of the screen, avoid speaking, and minimize eye-blinks.

Pre-processing analysis

The FieldTrip toolbox (https://www.fieldtriptoolbox.org)⁵² and Matlab version R2019b (The Mathworks, Natick, MA, USA) were used to analyze intracranial EEG data. By subtracting signals from nearby electrode contacts in the hippocampus, amygdala (Supplementary Fig. 1) or lateral cortex (Supplementary Fig. 2) recordings were converted for all participants to a bipolar derivation to optimize local activity^53–55. That is, we re-referenced data by subtracting activity from each contact from its adjacent contact within the same ROI.

If participants had more than one electrode in the amygdala or in the hippocampus, only the one located within the non-epileptogenic zone was used for the analysis. In the case of hippocampal electrodes, we exclusively included the most anterior ones. This was motivated by the stronger connectivity between amygdala and anterior, compared to the posterior, hippocampus⁵⁶ and is consistent with our previous study that investigated encoding-related amygdala and anterior hippocampus connectivity⁵.

For each amygdala and hippocampus channel, continuous recordings were divided into epochs from −7.5 to 7.5 s around stimulus onset. Given our focus on gamma responses, we did not apply any filtering for line noise within the gamma range. Trials were inspected visually in the time domain and time-frequency domain to identify artifacts produced by epileptic spikes or electrical noise. When artifacts were detected, the entire trial was discharged.

Spectral analysis

Each trial’s time-resolved spectrum decomposition was calculated using 7 Slepian multi-tapers for high frequencies from 35 to 150 Hz in 2.5 Hz steps. Slepian tapers were used with windows of 0.4 s width and a 10 Hz frequency smoothing. Low frequencies (<35 Hz) were calculated using single Hann taper, in 5 Hz steps and sliding windows were defined by 7 cycles per frequency step. The relative percentage change with respect to baseline (−1 to 0.1 s pre-stimulus time) was then calculated. Baseline corrected spectral activity was then averaged over patient trials and channels within the amygdala and the hippocampus.

Statistical analysis

Trials at retrieval were divided by emotion (aversive or neutral) and response (R, K, N). In line with previous work^4,5, KN responses were collapsed together as similar power spectral results were observed comparing Hit remembered trials with either KHit or Miss trials (Supplementary Fig. 3a–c). To investigate whether amygdala and hippocampal responses during recognition reflect aversive items reinstatement we restricted the analysis to previously seen scenes only. We used a cluster-based permutation test⁵⁷ in order to correct the family wise error rate in the context of multiple comparisons across time and frequency dimensions and identify significant interactions or main effects in the time-frequency domain. Using the maximum summation cluster statistic Montecarlo method, we conducted a 2-tailed t test. A paired t-test was performed at each time and each frequency bin for high (35–150 Hz) and low frequencies (0–34 Hz), separately, using a threshold of alpha = 0.025. Clusters were formed by temporal and frequency adjacency at each permutation step, selecting the a-priori time of interest in our data (from 0 to 1.5 s). Steps in the permutation were performed 10.000 times.

Global Encoding-Retrieval similarity (ERS) analysis

We quantified similarity of neuronal patterns in the gamma range (35–150 Hz) when the same item was presented at encoding and retrieval using representational similarity analysis (RSA)²⁴. We performed the same analysis in the amygdala and the hippocampus separately. Representational patterns were built concatenating epochs of time- and frequency-resolved power values (47 frequencies from 35 to 150 Hz in 2.5 steps) in windows of 0.5 s, overlapping by 0.4 s (80%)^10,25,26. A metric of similarity or reactivation was computed using Spearman´s correlation between encoding and retrieval from −0.5 to 1.5 s. Power data were baseline corrected from −1 to 0.1 s and if more than one bipolar channel was present, the average between channels was performed. For each trial, we defined a window of 0.5 s and included the time course of 35 to 150 Hz power (2.5 Hz steps). Our representational pattern thus consisted of a 2-dimensional matrix containing 47 (frequencies) and 50 (time points) values. To perform the correlations, we concatenated this matrix into a one-dimensional feature vector (of length 47 × 50 = 2350). To assess the temporal stability of this pattern we performed correlations across all possible combinations of time points within a trial. This resulted in an encoding x retrieval reactivation map for each trial (Fig. 3a), where the diagonal showed the reinstatement without lag between the two experimental phases and off-diagonal lagged correlations. We performed this map for each trial type (eRHit, eKHit&Miss, nRHit, nKHit&Miss) and tested for main effects and interactions using paired t-tests. We employed cluster-based permutation statistics to correct for multiple comparisons across time points⁵⁷. A paired t-test for each planned comparison was performed at each time point, clusters were formed by temporal adjacency with cluster threshold of alpha = 0.05 and a significant threshold of alpha = 0.025. Permutation steps were repeated 10,000 times.

Patterns locked to amygdala gamma peaks Encoding-Retrieval similarity (ERS) and Encoding-Encoding similarity (EES) analysis

Similar to the above analysis, we used RSA to quantify reactivation of neuronal activity but now selecting patterns of activity around amygdala high gamma peaks during encoding.

Amygdala gamma peaks selection

The time series of amygdala encoding data were band-pass filtered between 90 and 150 Hz using a two-pass finite impulse response (FIR) filter with a filter order of 3 cycles of the lower bound. We selected the high gamma range to avoid an overlap with the effect that we observed in the hippocampus at recognition (55–83 Hz). Amygdala gamma peaks at encoding were detected in the filter signal from 0.4 to 1.1 s. This is the window in which we observed a significant gamma effect for successful encoding of aversive scenes⁵. Only peaks with a minimum inter peak distance (IPD) of 0.1 s were considered. Each peak time was then used as a time-lock to determine frequency-resolved pattern of activity before and after the peak (±0.05 s).

Peak-locked ERS and EES analysis

We computed similarity between (1) amygdala at encoding and amygdala at retrieval; (2) amygdala at encoding and hippocampus at retrieval (ERS) and (3) amygdala and hippocampus at encoding (EES). In the ERS analysis, we did not synchronize amygdala or hippocampal activity with amygdala peaks, as the neurophysiological mechanisms involved in encoding and retrieval may differ in timing. Instead, we computed the ERS analysis 0.5 s before and 1.5 s after stimulus onset. By contrast, in the EES analysis, we aligned hippocampal data to the timing of amygdala peaks and computed EES from 0.5 s before until 1.5 s after the peaks. A schematic of the method can be found in Fig. 4a–d.

Both for the amygdala and the hippocampus, we used time-frequency baseline corrected data (see Method above) and only the most lateral amygdala and hippocampus channels in each patient were selected as previously reported for connectivity analyses between these structures⁵. For each frequency resolved pattern, the reactivation measure was calculated using Spearman correlation. To perform the similarity comparisons, we concatenated the representational patterns (25 [frequencies of gamma from 90 to 150 in 2.5 Hz steps] × 11 [0.1 s, ±0.05 s from the peak]) in a one-dimensional vector in each window. This was done in a temporally resolved manner proceeding in post-stimulus time steps of 0.1 s (25 × 11 = 275) and with an 80% overlap. In our EES analysis we locked hippocampal data to the time at which amygdala peaks occurred. In the ERS analysis, data was locked to the onset of stimulus presentation during encoding and retrieval. Single subject data were then averaged within each condition. Grand averages for each condition were calculated and we tested for the interaction (aversive remember (eRHit) – aversive known/forgotten (eKHit&Miss)) vs. (neutral remember (nRHit) – neutral known/forgotten (nKHit&Miss) applying a cluster-based permutation test⁵⁷.

To test whether the results were selective to the activity occurring only around amygdala peaks, we run the analysis using data not locked to the amygdala peaks. In each trial, we selected random segments of data of the same window length (0.1 s) and number as in the original analysis, ensuring that these segments were not locked to any identified peak and had a minim distance of 0.03 s from the peaks. We repeated this procedure 1000 times and assessed ERS and EES following the same procedure described in the original peak-locked analysis. Histogram of p-values for each random selection is reported in Supplementary Fig. 12.

To validate our findings, we also conducted a series of control analyses. Firstly, we repeated the test by increasing the minimum inter peak distance from 0.1 to 0.3 s to account for peaks autocorrelation, as reported in previous studies³⁴ (ERS: Supplementary Fig. 13a, EES: Supplementary Fig. 14a). Then, we randomly selected only one peak in each trial and repeated the analysis (ERS: Supplementary Fig 13b, EES: Supplementary Fig. 14b). All together, these control analyses allowed us to further validate the specificity of the observed effects and rule out any potential confounding factors related to peak-related data selection.

Hippocampus Encoding – Retrieval similarity: patterns at encoding were centered around the highest correlation coefficient of amygdala-hippocampus Encoding-Encoding similarity (EES)

We used RSA to quantify reactivation of neuronal activity in the hippocampus between encoding and retrieval. To achieve this, we focused on the moment of greater similarity between representational patterns in the amygdala and the hippocampus. Initially, we computed a correlation coefficient score over time for each amygdala pattern derived around the amygdala peaks at encoding. Then, for each observation we identified the time of greater similarity between the amygdala and hippocampal representational patterns by selecting the highest correlation coefficient. The delay between the timing of the amygdala gamma peak at encoding and the moment of the highest correlation score between the amygdala and hippocampus was then calculated. We constructed hippocampal representational patterns at encoding by considering activity at ±0.05 s from this event. As for the rest of ERS analysis, to perform the similarity comparisons we concatenated the representational patterns (25 [frequencies of gamma from 90 to 150 in 2.5 Hz steps] × 11 [0.1 s, ±0.05 s from the peak]) in a one-dimensional vector. Lastly, we computed the representational similarity between hippocampal patterns at encoding and hippocampal patterns at retrieval in a temporally resolved manner that proceeded from stimulus onset (0) until 1.5 s time steps of 0.1 s (25 × 11 = 275) and with an 80% overlap. A metric of similarity or reactivation was computed using Spearman correlation. Single subject data were then averaged within each condition. Grand averages for each condition were calculated and we tested for the interaction (aversive remember (eRHit) – aversive known/forgotten (eKHit&Miss)) vs. (neutral remember (nRHit) – neutral known/forgotten (nKHit&Miss) applying a cluster-based permutation test⁵⁷.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

B.A.S. designed the experiments and acquired the funding. B.A.S., M.C., and D.P.E. conceptualized the analysis. A.G.-N., R.T., L.I., and J.S. monitored patients and performed clinical evaluations. M.C. collected and analyzed the iEEG data. D.P.E. and B.A.S. provided input on the analysis and results. M.C., D.P.E., and B.A.S. wrote the manuscript, and all authors provided comments on the manuscript.

Peer review

Peer review information

Nature Communications thanks Liang Wang and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The electrophysiological data generated in this study have been deposited in UPM Drive under the following link: https://drive.upm.es/s/qq9kpFKmrdzAT9k. The raw patient data are protected and not available due to data privacy regulations and the conditions of ethical approval. Additional processed data supporting the findings of this study are available at the same link. Further information can be obtained by contacting the lead author.

Code availability

The MATLAB scripts and functions used to analyze the data in this study are publicly available via the GitHub repository https://github.com/mnlcosta/AversiveMemoryRetrieval, and have been archived on Zenodo with the following 10.5281/zenodo.15736783.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-61928-2.