Awake Hippocampal–Cortical Co-reactivation Is Associated with Forgetting

For several decades, research in rodents has demonstrated that neuronal firing patterns present at learning are replayed by hippocampal cells during post-encoding sleep (Girardeau & Zugaro, 2011; Skaggs & McNaughton, 1996; Wilson & McNaughton, 1993; Buzsaki, 1989) and awake rest (Jadhav, Kemere, German, & Frank, 2012; Davidson, Kloosterman, & Wilson, 2009; Karlsson & Frank, 2009; Diba & Buzsáki, 2007; Foster & Wilson, 2006), which preserve the spatiotemporal properties of previously learned representations. It is thought that neural reactivation, such as described in these studies, is a key memory consolidation mechanism that first acts to strengthen the encoding patterns within hippocampus (HPC; Carr, Karlsson, & Frank, 2012; Rasch & Born, 2007) and then gradually integrates new events with related representations stored in the cortex (Tambini & Davachi, 2019).

Systems consolidation views of memory, such as the complementary learning systems model, emphasize that consolidation occurs primarily through a coordinated communication between HPC and neocortex (Cowan, Schapiro, Dunsmoor, & Murty, 2021; Robin & Moscovitch, 2017; Kumaran, Hassabis, & McClelland, 2016; Winocur & Moscovitch, 2011; Nadel & Moscovitch, 1997; McClelland & O'Reilly, 1995), including sensory cortex. A small number of sleep studies in rodents have provided evidence for “coordinated replay” such that the HPC and category-selective cortices replay the same information around the same time (Lansink, Goltstein, Lankelma, McNaughton, & Pennartz, 2009; Wierzynski, Lubenov, Gu, & Siapas, 2009; Peyrache, Khamassi, Benchenane, Wiener, & Battaglia, 2009; Ji & Wilson, 2007).

Notably, in these types of models, the nature of representations across the HPC and cortex varies such that the HPC stores more veridical accounts of an event, whereas the cortex stores more schematic representations reflecting commonalities among similar events. These models leave open questions about how coordinated activation across the HPC and cortex relate to memory for unique events. Coordinated replay could strengthen detailed representations of the prior events, or alternatively, it could bias representations toward commonalities and/or introduce noise into the reactivated representation thereby degrading the precise details of encoded events.

Similar to what has been observed in rodents, an emerging body of work using human fMRI has reported post-encoding reactivation during awake rest (see Tambini & Davachi, 2019, for a review) and has shown that such reactivation appears to be related to better memory. More recently, researchers have looked at event-specific reactivation in the human brain, using multivariate approaches to examine the similarity between patterns of brain activity during encoding and post-encoding rest periods (Alm, Ngo, & Olson, 2019; Schlichting & Preston, 2014; Deuker et al., 2013; Staresina, Alink, Kriegeskorte, & Henson, 2013). For instance, Staresina et al. (2013) used representational similarity analysis (RSA) and found that human entorhinal cortex showed greater reactivation of subsequently remembered object–scene pairs during post-encoding rest. However, others have found that weakly encoded (Schapiro, McDevitt, Rogers, Mednick, & Norman, 2018) or weakly attended events (Jafarpour, Penny, Barnes, Knight, & Duzel, 2017) are prioritized for reactivation. Therefore, how awake reactivations in the human brain prioritize and consolidate information is still unclear, although common assumptions are made that these reactivations are beneficial for memory.

Notably, none of this prior work investigating event-specific reinstatement probed the role of sensory representations in conjunction with the HPC, precluding the ability to evaluate the consequence of processes in line with systems consolidation. The question of whether or how reactivations are coordinated between the HPC and cortex has not been directly addressed in human fMRI studies. Studies that investigated hippocampal–cortical interactions during awake post-encoding rest utilized connectivity as a proxy for “coordinated replay” and reported greater functional connectivity between the HPC and areas of the cortex during post- compared with pre-encoding rest for subsequently remembered information (Liu, Grady, & Moscovitch, 2018; Murty, Tompary, Adcock, & Davachi, 2017; Tompary & Davachi, 2017; Gruber, Ritchey, Wang, Doss, & Ranganath, 2016; Tompary, Duncan, & Davachi, 2015; Tambini, Ketz, & Davachi, 2010). To our knowledge, no human fMRI study has yet tested event-specific hippocampal–cortical coordinated reactivations (co-reactivations).

Here, we tested two hypotheses, derived from theories of systems consolidation: (1) The HPC and cortex should co-reactivate information, and (2) this co-reactivation should correlate with subsequent memory performance. To assess these hypotheses in a more targeted way, we investigated the occurrence of coordinated reactivation of encoding patterns occurring in the HPC and cortex during post-encoding awake rest. We expanded on methods established in prior work (e.g., Schapiro et al., 2018; Gruber et al., 2016; Staresina et al., 2013) to look for co-reactivation on a repetition time (TR)-by-TR basis, providing a proof of concept that concurrent reactivation patterns across brain regions can be identified during rest periods using fMRI. To test this, while in the scanner, participants first completed a 6-min rest scan (pre-encoding), followed by two encoding blocks, which were each followed by a posttask rest scan (post-encoding; Figure 1A). During encoding, participants were shown images of one of two possible categories (face or scene, in separate blocks) paired with a unique location on a 4 × 4 grid (image trials, henceforth). Leveraging our design, our analyses focused on examining co-reactivation during the post-encoding rest periods between the HPC and two cortical ROIs in the fusiform face area (FFA) and parahippocampal place area (PPA). Cortical ROIs were a priori selected given their previously established role in processing category-selective information (e.g., faces in FFA, Kanwisher & Yovel, 2006; and scenes in PPA, Park & Chun, 2009; Epstein & Kanwisher, 1998). We defined co-reactivation as reactivation in HPC and cortex to the same image trial, at the same TR (± 1 TR; see Figure 1B and C). We then examined the relationship between co-reactivations and subsequent memory based on recognition tasks conducted immediately upon leaving the scanner, 1-day, and 1-week after learning. On these memory tasks, participants were cued with a location on a 4 × 4 grid and were instructed to select the correct image paired with this location from among three image options (a target and two lures). Our analyses revealed that there were more hippocampal–cortical co-reactivations during post- than pre-encoding rest, establishing this novel TR-based co-reactivation approach as a successful method to study hippocampal–cortical interactions during the consolidation window. Moreover, we found that these hippocampal–cortical co-reactivations were uniquely associated with forgetting, posing questions about their functional role.

Thirty-seven right-handed young adults (M_age = 22.1 years, SD = 3.16 years, 20 women) from Temple University and the surrounding community participated in the study between 2019 and 2022. All participants spoke English and were free of neuropsychiatric disorders and MRI contraindications. Data collection was disrupted because of the COVID-19 pandemic, resulting in two cohorts of 21 participants and 15 participants, respectively. One participant from the first cohort and two participants from the second cohort were excluded from the analyses because of little-to-no event-related MRI signal. Comparisons between the cohorts regarding their behavioral memory performances and imaging quality did not reveal any significant differences; therefore, we collapsed the data across the two cohorts for all the reported analyses, resulting in a sample of 34 participants, which is higher than reported sample sizes in majority of the previous research investigating related phenomenon (i.e., reactivation; see Table 1).

Participants completed three rest and two task (encoding) runs (Figure 1A). Scanning started with a baseline (pre-encoding henceforth) rest run (6 min), which was used as a control rest period for reactivation and co-reactivation analyses (see below, Co-reactivation Analysis section). Participants then completed two encoding task runs (8 min each), each of which was followed by a post-encoding rest period (6 min each). The order of encoding blocks (face-first or scene-first) was counterbalanced across participants. Total fMRI scan time was 34 min.

The task used an event-related design. Each trial began with a fixation cross (1 sec) followed by a face or scene (4 sec) located in a specific location on a 4 × 4 grid (“image trial”). This was followed by a 10-sec long odd-or-even number-judgment task during which time they were shown random numbers (2 sec each) and were asked to press a button to indicate whether it was an odd or even number (referred to as “number trials”). The purpose of the number trials was to provide time to allow us to model the fMRI signal for each image trial while imposing a task that dampened overt rehearsal. Total trial length was 16 sec. There were eight images per face/scene category, and each image was repeated 4 times in total. Face and scene images were paired with different parts of the grid to avoid any overlap between the paired associates across categories, and their distribution across the four quadrants of the grid was also counterbalanced. Participants were instructed to pay attention to the image trials, to press a button on the number trials, and to keep their eyes open during the rest scans.

Upon exiting the scanner, memory was tested immediately, 1-day later, and 1-week later, with two surprise tasks: a cued-location recall and a cued-image recognition task. Cued recall and recognition were completed in separate blocks, and their order was counter-balanced across participants. In this article, we focus on the recognition task because the cued-location recall task showed susceptibility for “new learning” occurring during the repeated testing contexts, such that this measure did not show forgetting over delays. During recognition, participants were shown a 4 × 4 grid, with one of the cell locations highlighted with black, together with three image options (one target and two old lures from the same category, which were also previously encoded). They were instructed to “choose (click on with the mouse) the correct image that had been paired with the highlighted location during encoding.” There were 16 recognition trials, evenly split between faces and scenes. A categorical accuracy variable (correct/incorrect) was derived for each trial, and a total correct variable was created across all the trials for each participant.

Using separate chi-square tests, we first tested the frequency of correct versus incorrect trials for recognition for each test day (Figure 2). We then examined whether there were significant changes in trial-based recognition accuracy across the three test days. To this end, we first created three 2 × 2 contingency tables with the trial counts for behavior change from immediate to 1-day delay, from immediate to 1-week delay, and from 1-day to 1-week delay conditions. For these tables, behavior change was coded as “correct at both tests,” “initially correct – then incorrect,” “initially incorrect – then correct,” and “incorrect at both tests.” Because these are paired data, we then conducted three separate McNemar's tests (McNemar, 1947) for the behavioral change across days.

MRI scans were completed at Temple University on a 3-T Siemens Magnetom Prisma scanner, using a 64-channel phased-array head coil. High-resolution T1-weighted (T1w) anatomical images were collected using a three-dimensional magnetization prepared rapid acquisition gradient echo pulse sequence (TR = 520 msec, echo time = 0.007 msec, field of view = 100 mm, flip angle = 60°, 2-mm slice thickness). Functional T2*-weighted images were collected using a gradient-echo planar pulse sequence with the following parameters: TR = 2000 msec, echo time = 29 msec, field of view = 100 mm, flip angle = 76°, 2-mm slice thickness. DICOM (Digital Imaging and Communications in Medicine) images were converted to NIFTI (Neuroimaging Informatics Technology Initiative) format with Brain Imaging Data Structure nomenclature using dcm2niix (Gorgolewski et al., 2016). Quality control was achieved by running the MRIQC pipeline (Version 0.10.4 in a Docker container; Esteban et al., 2017) on the structural and functional images.

fMRI preprocessing was performed with FMRIB Software Library (FSL) 6.0.1. (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012). First, the T1w anatomical image was skull stripped using the Brain Extraction Tool. This image was used to assist in spatial normalization processes detailed below. Brain tissue segmentation of white matter (WM), gray matter, and cerebrospinal fluid (CSF) was performed on the brain-extracted T1w images using FAST (FMRIB's Automated Segmentation Tool). These segmentations were used to extract time series from the WM and CSF for reduction of noise in our preprocessing stream. fMRI preprocessing was completed using the fMRI Expert Analysis Tool (FEAT) version as implemented in FSL 6.0.1. using a pipeline designed to minimize the effects of head motion (Murty et al., 2018). This included simultaneous head motion correction and non-linear warping to the Montreal Neurological Institute (MNI) space, but no temporal or spatial filtering. The same preprocessing methods were applied to all encoding and rest runs.

HPC, FFA, and PPA were a priori selected for their role in memory and category-selective processing, respectively, and were functionally defined in the MNI space at the group level. We first extracted univariate activity for each encoding task and then completed a higher-level statistical analysis in FEAT version as implemented in FSL 6.0.1. to extract a group average of face > scene and scene > face contrasts. Active clusters from the face > scene contrast was used to extract the peak coordinates for FFA (MNI: left: x = 64, y = 43, z = 24; right: x = 23, y = 41, z = 25), whereas active clusters from the scene > face contrast was used to extract the peak coordinates for PPA (MNI: left: x = 57, y = 37, z = 31; right: x = 29, y = 44, z = 29), thresholding the z-stat maps at 3.1. Using a lower threshold (1.5), the peak hippocampal voxels that were active in both face > scene and scene > face contrasts were identified for HPC (MNI: left: x = 56, y = 47, z = 33; right: x = 34, y = 50, z = 32). For each ROI, we created a sphere using a 5-mm radius kernel with the fslstats command in FSL around the peak coordinates. Importantly, each mask was created in MNI space and binarized before extracting any activity patterns from the task or rest scans.

After preprocessing, we ran two separate general linear models (GLMs) for face-location and scene-location blocks, which modeled each image trial as a separate regressor. Importantly, each event regressor included all four repetitions of that image trial to increase detection power. Therefore, each GLM included eight event regressors, each modeled for a 4-sec duration and were convolved with a double-gamma hemodynamic response function. Six head-motion parameters, and their first derivatives, and time series extracted from CSF and WM were added as covariates to the model to reduce noise, and a 2-mm FWHM kernel was used for spatial smoothing. Voxel-vise encoding activity was extracted from the t-stat maps for each image trial, within each ROI. We chose to use t-statistics because doing so addresses the noise from highly variable voxels (Dimsdale-Zucker & Ranganath, 2018; Misaki, Kim, Bandettini, & Kriegeskorte, 2010).

All GLMs were run using FEAT Version 6.0 as implemented in FSL 6.0.3. First-level face > baseline and scene > baseline contrasts were estimated in our ROIs, separately for each hemisphere (see next section for ROI selection). Finally, we modeled all three rest scans (i.e., the pre-encoding and two post-encoding rests) in GLMs with the same nuisance parameters. We then obtained and high-pass filtered the residuals from these models and extracted TR-based activity from the residual t-stat maps.

First, reactivation of encoding events was quantified using RSA (Kriegeskorte, Mur, & Bandettini, 2008). Within each encoding session for each participant, each trial's encoding patterns were correlated (Pearson's correlation coefficients) to the pattern of activity at each TR of the rest periods to find potential reactivations. Importantly, we completed this analysis in both pre- and post-encoding rest periods, which resulted in four (face-pre-encoding, face-post-encoding, scene-pre-encoding, and scene-post-encoding) 8 (encoding trials) × 180 (rest TRs) reactivation matrices (see Figure 1B for an example) for each ROI. All correlations were then Fisher z-transformed.

Theoretically, there should not be any “reactivations” during pre-encoding rest, given that the participants did not see any of the image trials before entry into the scanner. That is, patterned activity during the pre-encoding rest phase represents true baseline activity, and therefore, any correlation between pre-encoding rest activity and an event's encoding activity would necessarily be a spurious one. Thus, in line with previous work (e.g., Gruber et al., 2016; Staresina et al., 2013), we reasoned that the average correlation between an event's encoding activity and pre-encoding rest activity could define a data-driven cutoff for spurious correlations, and thus can be used for thresholding the reactivation patterns that were detected at post-encoding: By extracting encoding activity (i.e., beta values) for each event and then correlating these values with each pre-encoding TR, we could establish the frequency and value range for all observed correlations. Using those values, we calculated the average similarity of an event to the pre-encoding rest and then defined a cutoff threshold at 1.5 SDs above this mean (1.5 value based on prior work by Staresina et al. [2013] and Schapiro et al. [2018]). This threshold was then applied to the post-encoding data to minimize the contribution of any spurious correlations for the given event at post-encoding rest. That is, only the post-encoding TRs that survived each image's calculated threshold were considered potential “reactivations” for that image trial (Figure 1B). Next, we counted the number of such reactivating TRs to define our post-encoding reactivation count variable. This analysis was repeated for each event in all our ROIs, and separately using post-face rest for face-location trials, and post-scene rest for scene-location trials.

We next counted TRs (± 1) that reactivated the same image trial across our ROIs. TRs that reactivated the same image trial across one of the HPCs (left or right) and one of the cortical regions (left or right FFA [or PPA]) were counted as co-reactivating TRs for two ROI pairs: HPC–FFA and HPC-PPA (Figure 1C). All reactivation and co-reactivation analyses were completed in a custom MATLAB code (Version 2020b, available at https://www.mathworks.com/products/new_products/release2020b.html).

Importantly, our a priori hypothesis was that FFA should reactivate faces (Kanwisher & Yovel, 2006) and PPA should reactivate scenes (Park & Chun, 2009; Epstein & Kanwisher, 1998) given their suggested role in category-selective processing. However, we did not find any significant differences between FFA and PPA in selectively reactivating faces and scenes, respectively; therefore, we collapsed trials across categories for all reactivation and co-reactivation analyses. This collapsed data included post-face rest for face trials, and post-scene rest for scene trials as their respective post-encoding rest in all the reported analyses.

Finally, while the main analyses utilized the described (co-)reactivation count variables (i.e., frequency of (co)reactivations), we also extracted a (co-)reactivation strength variable to support additional analyses enabling comparison of our findings to previously reported studies that focused on reactivation strength (Schapiro et al., 2018; Staresina et al., 2013). Specifically, we calculated the sum of the Pearson's r values for all reactivating TRs for each image trial, within each region (i.e., region-specific reactivation strength). We then calculated the sum of the Pearson's r values for all co-reactivating TRs (separately for HPC–FFA and HPC-PPA) for each image trial (i.e., co-reactivation strength). Main analyses that were conducted initially using the (co-)reactivation count variables were then repeated using these (co-)reactivation sum variables.

Using item-level multilevel linear models, entering the subject information as random slopes, we tested whether reactivation counts were predicted by rest period (post-encoding vs. pre-encoding) and by hemisphere (left vs. right), separately for each ROI. All models included a Rest × Hemisphere interaction term to test whether there were any lateralization effects for the reactivation counts. Results were Bonferroni-corrected for multiple comparisons at p_adjusted = .017. For the ROIs that showed significantly higher reactivation at post-encoding than pre-encoding rest, we then tested whether these reactivations were associated with recognition accuracy (Figure 3). For ROIs that showed a lateralization effect (i.e., only one hemisphere had significantly higher post-encoding reactivation counts), we have tested the memory associations within the hemisphere that showed the significant lateralization effect. For ROIs without such lateralization, we collapsed the reactivation counts across hemispheres before testing the associations with recognition memory.

We next tested two item-level multilevel regression models to predict co-reactivation counts from the rest periods for each ROI pair (HPC–FFA and HPC-PPA; Bonferroni corrected at p_adjusted = .025). Using these results as a filter (e.g., retaining the ROI pairs that showed significantly higher co-reactivation at post-encoding than pre-encoding rest), we then tested whether the co-reactivation counts were significantly related to memory recognition (Figure 4).

The unstandardized beta coefficients are reported for all our significant results. Reported statistical analyses were performed using R software (R package Version 3.4.1) using mcnemar.test (rcompanion library), the cor.test, t.test, aov, lm (the stats library), and lmer (the lme4 library) functions depending on the test. All continuous variables were standardized before testing the regression models. Analysis scripts are available upon request.

Using separate chi-square tests, we first tested the frequency of correct versus incorrect responses for recognition for each test day. These tests revealed that across all test days, participants had significantly higher numbers of correct than incorrect responses (Immediate: χ²(1) = 256, p < .0001; 1-day delay: χ²(1) = 219, p < .0001; 1-week delay: χ²(1) = 159, p < .0001). Given that participants repeated the same recognition memory test, we next asked how their accuracy changed across days for image trials. Using McNemar's test, we found that there were no significant changes in behavior from immediate to 1-day delay condition (p = .2), or from 1-day to 1-week delay (p = .2) conditions. Importantly, however, participants' responses indicated significant forgetting from the immediate to 1-week delay test (McNemar's χ²(1) = 9, p = .003).

Using RSA, we first identified pre- and post-encoding TRs that showed significant reactivations of each image trial in our a priori ROIs (Figure 1B). Our category-specific tests showed nondifferentiated reactivation counts in FFA and PPA for faces and scenes, and we collapsed across categories for the following reactivation and co-reactivation analyses.

We first tested item-level multilevel linear regression models, separate for each ROI, to determine whether reactivation counts differ across pre- and post-encoding rest and across hemispheres, thereby including a Rest (pre-/post-encoding) × Hemisphere (left/right) interaction term. For HPC, we found a main effect of Rest (post- > pre-encoding: β = 0.9, SE = 0.39, p = .02), but not a main effect of Hemisphere (β = 0.04, SE = 0.39, p = .9), or a significant Rest × Hemisphere interaction (β = 1.03, SE = 0.55, p = .06). Pairwise comparisons revealed that both right and left HPC showed increased number of reactivations at post- than pre-encoding rest, although effects were more prominent in right HPC (left: β = 0.9, SE = 0.39, p = .09; right: β = 1.93, SE = 0.39, p < .001).

In FFA, we found a trending effect of Rest (post- > pre-encoding: β = 0.76, SE = 0.45, p = .09) and no significant main effect of Hemisphere (right > left: β = 0.39, SE = 0.45, p = .4). However, there was a significant Rest × Hemisphere interaction (β = 2.21, SE = 0.64, p = .001). Post hoc pairwise comparisons revealed that there were greater number of reactivations during the post- than pre-encoding rest in the right FFA (β = 2.97, SE = 0.45, p < .001), but not in the left FFA (β = 0.76, SE = 0.45, p = .33). For PPA, there was a main effect of Rest, with a greater number of reactivations during post- than pre-encoding rest (β = 0.79, SE = 0.36, p = .027). However, there was neither a significant main effect of Hemisphere (β = 0.17, SE = 0.36, p = .66) nor a significant Rest × Hemisphere interaction (β = −0.17, SE = 0.51, p = .76).

Next, we asked whether event-specific post-encoding reactivations were associated with subsequent recognition memory test performance in regions that showed an increased reactivation count at post-encoding rest (i.e., right FFA, bilateral HPC; see Methods and SI sections for details). Interestingly, we found that there were a greater number of reactivations in the right FFA for trials that were incorrectly recognized on the 1-week delayed recognition task compared with correctly recognized image trials, F(34, 1024) = 2.9, β = 3.27, SE = 1.13, p = .004 (Figure 3). Importantly, the number of reactivations in the right FFA during the pre-encoding rest period did not differ based on subsequent memory performance, F(34, 1024) = −0.99, β = 0.49, SE = 0.5, p = .33, indicating that the phenomenon is experience dependent. No other ROIs showed significant differences in reactivation counts based on subsequent recognition performance after a 1-week delay.

The prior analyses looked at one region in isolation, but as detailed above, our main goal was to examine coordinated reactivation across the HPC and cortex in line with processes of systems-like consolidation. Thus, we next quantified the TRs that showed significant reactivation to the same image trials in HPC and at least one cortical ROI (e.g., FFA or PPA; Figure 1C), creating a co-reactivation count variable between bilateral HPC-bilateral FFA and bilateral HPC-bilateral PPA. Using item-level multilevel linear modeling, we tested a model to predict co-reactivation counts from the post- and pre-encoding rest periods, separately for each of the hippocampal–cortical ROI pairs. First, examining co-reactivation between HPC and FFA, we found significantly higher co-reactivation counts during the post-encoding compared with pre-encoding rest periods, F(34, 1024) = 4.33, β = 0.79, SE = 0.18, p < .001. There was no difference in counts for HPC and PPA, F(34, 1024) = 1.7, β = 0.3, SE = 0.17, p = .08; therefore, we focused the remaining analyses on HPC–FFA co-reactivations.

We next tested whether the HPC–FFA co-reactivation counts during post-encoding rest were associated with recognition performance. This analysis revealed that there were higher HPC–FFA co-reactivations counts for incorrectly recognized images at a 1-week delay, F(32, 480) = 2.89, β = 1.18, SE = 0.41, p = .004, but not the earlier memory tests (Figure 4). Importantly, there was no significant difference in subsequent memory for items as a function of “co-reactivation” during the pre-encoding rest period (p = .67), further suggesting that the effects observed at post-encoding rest are experience dependent. Considering that the within-region reactivation effects (post > pre) were more evident in the right than left hemisphere, for both HPC and FFA, we additionally tested whether these co-reactivation effects might be especially right lateralized. To test this, we recounted HPC–FFA co-reactivations separately for the right and left hemispheres, and tested a Hemisphere × Rest interaction model for HPC–FFA co-reactivation counts. The results showed that there was a significant Hemisphere × Rest interaction (β = 0.32, SE = 0.11, p = .004). Pairwise comparisons revealed that there was a significantly higher count for HPC–FFA co-reactivations in post- than pre-encoding rest in the right hemisphere (β = 0.25, SE = 0.08, p = .007), and these co-reactivations were significantly associated with incorrectly recognized image trials at 1-week delay (β = 0.36, SE = 0.17, p = .03).

The above analyses, showing a negative relationship between co-reactivations and memory at a 1-week delay, suggest that the coordinated reactivation occurring during post-encoding rest might actually impair associative recognition. Importantly, we did not show this significant inverse relationship for the immediate and 1-day delayed tests, suggesting that this relationship between reactivation and memory may reflect processes involved in memory consolidation. A more specific way to test whether this relationship is related to consolidation is to use a measure of forgetting, in which we compare changes in memory across tests. Accordingly, we coded images based on their relative status at the immediate and delayed tests. Image trials that were correct at immediate and remained correct at the 1-day or 1-week delayed test were coded as “remembered.” Image trials that were initially correct at immediate but were incorrect at the 1-day or 1-week delayed test were coded as “forgotten.” We then tested whether there was any significant relationship between HPC–FFA co-reactivations and remembered and forgotten trials, tested separately for changes from immediate-to-1-day delayed, and immediate-to-1-week delayed. We found a consistent pattern of relationship for forgetting after 1 day and forgetting after 1 week. Specifically, we found that trials that were initially correct at immediate but then were forgotten at 1-day delay showed a greater number of post-encoding HPC–FFA co-reactivations, F(32, 414) = 2.03, β = 1.09, SE = 0.54, p = .043. Similarly, we found that trials that were initially correct at immediate but then were forgotten at the 1-week-delay condition showed a greater number of post-encoding HPC–FFA co-reactivations, F(32, 407) = 2.16, β = 1.02, SE = 0.47, p = .031 (Figure 5). We did not find any significant differences in the number of HPC–FFA co-reactivations for remembered and forgotten trials when we tested the model for changes from 1-day to 1-week delay testing, F(31, 386) = 1.58, β = 0.78, SE = 0.5, p = .11. Together, these findings suggest that post-encoding HPC–FFA co-reactivations were uniquely associated with forgetting over time.

Given our initial observation that greater reactivation counts in right FFA were significantly associated with incorrectly recognized image trials at 1-week delay, we asked whether the association between HPC–FFA co-reactivations and incorrect delayed recognition performance was driven by the reactivations in the right FFA alone. To this end, we examined TRs in the right FFA that co-reactivated a given image with HPC, which we refer to as “coordinated,” versus right FFA TRs that did not show any co-reactivations (“uncoordinated”). Using the total number of coordinated and uncoordinated TRs in right FFA, we then retested our original model for the right FFA and 1-week-delayed recognition association. The results revealed that the observed relationship was specific to coordinated reactivations. Namely, the number of coordinated TRs, but not uncoordinated TRs, in the right FFA was greater for incorrectly recognized compared with correctly recognized trials at 1-week delay (coordinated: F(32, 480) = 2.48, β = 1.16, SE = 0.47, p = .014; uncoordinated: F(32, 480) = 1.82, β = 1.21, SE = 0.67, p = .07; Figure 6). Furthermore, we did not find any significant relationship between 1-week delayed-recognition and uncoordinated TRs in HPC (uncoordinated in left HPC: F(32, 480) = 0.16, β = 0.08, SE = 0.53, p = .88; uncoordinated in right HPC: F(32, 480) = 1.36, β = 0.76, SE = 0.56, p = .18). Together, these findings support the conclusion that it is the coordination of the HPC and FFA reactivations that is predictive of poorer subsequent recognition, rather than merely the reactivation of an individual region. In other words, when these regions reactivate independently, they do not show any significant relationship with subsequent memory. However, when they co-reactivate, they are associated with forgetting.

Our main analyses reported above focused on the total number of reactivations during post-encoding rest, but previous work has utilized other measures of awake reactivations, including the sum (Schapiro et al., 2018) or mean (Staresina et al., 2013). Interestingly, unlike the present results, these prior studies have shown that post-encoding reactivation is related to better subsequent memory, rather than forgetting. As such, we next asked whether our unique finding is driven by this distinction in methods. To test this, we generated a sum of reactivation scores for each unique event, in line with the method outlined by Schapiro et al. (2018). Briefly, for each image, we calculated the sum of the Pearson's r values across all the TRs showing reactivation above our threshold (see Methods section). Thus, instead of generating a count of the number of reactivations or co-reactivations present across the rest periods, the sum of reactivations measures the strength of the correlation between the pattern at encoding and rest. We then tested the same interaction models reported above, using this event-specific reactivation sum score instead of reactivation counts. In HPC, we found a significant main effect of Rest (β = 0.31, SE = 0.13, p = .021), showing greater sum of reactivations during post- than pre-encoding rest period. Whereas the main effect of Hemisphere was trending (β = 0.23, SE = 0.13, p = .08), the Rest × Hemisphere interaction was significant (β = 0.39, SE = 0.18, p = .037). Pairwise comparisons revealed that the reactivation sum was greater during post-encoding versus pre-encoding rest in right HPC (β = 0.7, SE = 0.13, p < .0001) but not in the left HPC (β = 0.7, SE = 0.13, p = .1). In the FFA, this analysis did not yield a significant main effect of Rest (β = 0.12, SE = 0.2, p = .55) or hemisphere (β = 0.28, SE = 0.2, p = .16), but there was a significant Rest × Hemisphere interaction (β = 1.14, SE = 0.28, p < .001). Pairwise comparisons revealed that the right FFA showed greater reactivation sum during post- than pre-encoding rest (β = 1.26, SE = 0.2, p < .001), but not in the left FFA (β = 0.1, SE = 0.2, p = .9). Finally, in the PPA, we found only a trending level main effect of Rest (post- > pre-encoding: β = 0.28, SE = 0.16, p = .08), but a significant main effect of Hemisphere (β = 0.33, SE = 0.16, p = .043). There was no significant Rest × Hemisphere interaction in PPA (β = 0.01, SE = 0.23, p = .96).

Next, we asked whether the event-specific post-encoding reactivation sums are associated with recognition performance in these regions showing a significant increase in reactivation sums (right HPC and right FFA). We did not find any significant reactivation sum and recognition memory effects in the right HPC (β = 0.46, SE = 0.32, p = .15). We did, however, replicate our original effect in the right FFA (β = 1.47, SE = 4.94, p = .003) for incorrectly recognized image trials at 1-week delayed recognition. Importantly, this heightened reactivation sum − incorrect recognition association was not significant at the pre-encoding rest period (p = .34).

Turning next to the measures of co-reactivations, we examined if our results differed when examining the sum, rather than total number, of co-reactivations between HPC and cortex. As above, we calculated the sum of the correlation values over all the TRs showing above threshold co-reactivation for a given pair. Replicating our findings with co-reactivation counts, we found significantly greater sum of co-reactivations between HPC–FFA during post-encoding rest compared with pre-encoding rest, F(34, 1024) = 4.2, β = 0.61, SE = 1.46, p < .0001. Moreover, this increased HPC–FFA co-reactivation sum was also associated with incorrect recognition at the 1-week-delayed test condition, F(32, 480) = 2.64, β = 0.86, SE = 0.32, p = .008. Importantly, the co-reactivation sum at pre-encoding rest for HPC–FFA did not significantly differ based on recognition (p = .68). Finally, for the HPC–PPA co-reactivation sum, we did not find any significant differences between post- and pre-encoding rest (p = .1). As such, we did not calculate any follow-up tests on the relationship with memory.

Together, we replicated all significant effects that relied on reactivation counts when using reactivation sums, including the relationship with incorrect recognition. Thus, the differences between our findings and that of previous work (e.g., Schapiro et al., 2018; Staresina et al., 2013) cannot be attributed simply to differences in methodology.

For the last 2 decades, researchers have shown that, in rodents, hippocampal reactivations during sleep or awake rest recapitulated previous events, and in some cases, correlates with better subsequent memory (Jadhav et al., 2012; Girerdeau & Zugaro, 2011; Davidson et al., 2009; Karlsson & Frank, 2009; Diba & Buzsáki, 2007; Foster & Wilson, 2006; Skaggs & McNaughton, 1996; Wilson & McNaughton, 1993; Buzsaki, 1989). More recently, this work has been extended to show that the HPC and cortex can also co-replay previously encountered locations (Lansink et al., 2009; Peyrache et al., 2009; Wierzynski et al., 2009; Ji & Wilson, 2007).

Here, we asked whether in humans a similar sort of co-replay of individual events occurs, and how it affects subsequent memory. Extending prior work, we provide novel evidence specifically examining coordinated reactivation between HPC and cortex during periods of post-encoding awake rest. First, we showed that there was an increased number of co-reactivations between the HPC and FFA for individual experienced events during post-encoding (as compared with pre-encoding) awake rest. Moreover, in addition to significantly higher frequency of co-reactivations at post- than pre-encoding, the analysis of our data using a measure indexing the strength of these co-reactivations showed the same results. Importantly, the appearance of increased HPC–FFA co-reactivations during the post-encoding rest period strongly indicates that these co-reactivation patterns reflect experience-dependent changes related to encoded events, and not just spurious similarities. Second, we found that HPC–FFA co-reactivations were associated with diminished memory performance at 1-week delay, meaning, trials that showed greater coordinated reactivation across regions were incorrectly recognized on the delayed memory test. Finally, we found that these co-reactivations were specifically associated with the forgetting of trials that had been initially correctly recognized at immediate testing, suggesting that the observed phenomenon is not merely the result of failed encoding, but rather is linked to post-encoding consolidation processes that predict the later weakening or blurring of representations.

Some prior fMRI work has shown that item-specific post-encoding reactivations were associated with better subsequent memory (e.g., Deuker et al., 2013; Staresina et al., 2013). As such, the seemingly paradoxical findings from the present evaluation of how co-reactivations relate to memory raise important questions about exactly how hippocampal–cortical co-reactivations may transform event memories in addition to, or differently from, regionally limited reactivations as they have been more traditionally assessed.

Systems consolidation theories posit that hippocampal and cortical representations are different in nature: The HPC is posited to be engaged in pattern separation (Yassa & Stark, 2011), supporting storage of more detailed episodic representations, whereas the cortex is theorized to generalize and store a more gist-level representations of events (Moscovitch & Gilboa, 2021; Kumaran et al., 2016). Importantly, these different representations can coexist and interact with one another, depending on the task demands (Moscovitch & Gilboa, 2021; Sekeres, Winocur, & Moscovitch, 2018; Winocur & Moscovitch, 2011). This interaction is likely reflected in hippocampal–cortical co-reactivations. However, any expected impacts of co-reactivations on subsequent memory are not fully fleshed out in current memory consolidation theories. One possibility is that these different representations could bias the memory decision in opposing directions: Whereas cortical representations might inform gist-dependent memory judgments, hippocampal representations might bias memory judgment toward unique details. Any conflict between the gist representation with the specific event representation could lead to interference and, thus, forgetting. We speculate that our central observation—that co-reactivations relate to incorrect recognition and forgetting—may be explained by conflicts that arise when the HPC and cortex attempt to represent and communicate different versions of events; with hippocampal reactivations pushing toward pattern separation among image representations, but cortical reactivations evincing more gist-like memory representations.

Another possibility regarding the interaction of coexisting hippocampal and cortical representations is that it reflects cross-regional interactions that lead to the integration of related events, and that this integration itself may disrupt subsequent representation of event-specific encoding patterns. This interpretation would be more in line with trace transformation theories of consolidation, which proposed that memories undergo re-organization during consolidation, through which they are transformed into more generalized variants of themselves, retaining gist but losing details and context specificity (Moscovitch & Gilboa, 2021; Sekeres et al., 2018). Consistent with this idea, Tompary and Davachi (2017) showed that events that share overlapping associations become increasingly similar in their cortical (in this case medial prefrontal cortex [mPFC]) representations over time. Moreover, this increased similarity for overlapping events was associated with increased HPC–mPFC functional connectivity during post-encoding rest, providing further evidence that post-encoding processes help prioritize integration of related events through their commonalities. As stated above, this prioritization of common features during consolidation processes may result in forgetting of unique features that would otherwise enable recollection of individual events. Evidently, when such forgetting occurs, gist-level schematic knowledge influences memory retrieval more (Tompary, Zhou, & Davachi, 2020). This account provides an alternative interpretation for why we find higher HPC–FFA co-reactivations for trials that are incorrectly recognized at 1-week delay. It is likely that these co-reactivations have blurred out the unique features of these events, thereby making it difficult to distinguish the target image from the two other alternative lure images, which were themselves also studied during encoding. Indeed, using a small number of grid locations as the paired associates to these images may have biased the task toward producing such generalization, especially because the grid locations on their own do not provide any discriminable features that could help to uniquely identify individual events. This state of affairs might be different in important ways from previous tasks in which reactivated items were paired with common associates of the items (e.g., pairing a cat image with a meow sound; Oudiette, Antony, Creery, & Paller, 2013).

We did not find any significant relationship between HPC–FFA co-reactivations and recognition at immediate or 1-day delay testing. Although at first it appears inconsistent with previous research (e.g., Gruber et al., 2016; Tambini et al., 2010) in that we did not observe any (co)-reactivation effects on immediate recognition performance, it is important to note that behaviorally, recognition accuracy was very high at immediate test, highlighting successful learning, perhaps because of the fact that stimuli were repeated during encoding. That is, the lack of variance in immediate recognition performance may have contributed to the lack of relationship with HPC–FFA co-reactivations. Moreover, the consolidation effects that we observe later are in the direction of forgetting. Therefore, the lack of a significant relationship between HPC–FFA co-reactivations and immediate recognition is not unexpected in our unique data. It is puzzling, however, that we did not find a significant relationship between HPC–FFA co-reactivations and recognition, when comparing correct versus incorrect responses, at the 1-day delay condition, as commonly observed by others (e.g., Alm et al., 2019; Schapiro et al., 2018). However, it is important to highlight that we found a significant relationship between HPC–FFA co-reactivations and the memory change from immediate to 1-day delayed recognition (Figure 5A) such that images that were forgotten overnight were more frequently co-reactivated during post-encoding rest. These results, together with the similar effects that emerged after a 1-week delay, are in line with the hypotheses that memories are transformed over time, with consolidation (Moscovitch & Gilboa, 2021; Cowan et al., 2021; Kumaran et al., 2016), and that gist-level, schematic knowledge influences memory retrieval more after forgetting of episodic details (Tompary et al., 2020).

So far, we have offered several alternative interpretations, aligning with different assumptions from systems consolidation theories, for our finding that HPC–FFA co-reactivations are associated with incorrect recognition. However, it is also possible that the mechanisms underlying our findings may be better explained by theoretical frameworks distinct from consolidation theory. Here, we consider an important question regarding the function of awake (co-)reactivations: Does awake spontaneous post-encoding reactivation always benefit memory?

Awake reactivations have been considered highly similar to reactivations during sleep (Tambini & Davachi, 2019), yet it is likely that the outcomes of reactivation processes differ in different mental states. One study provided early evidence that awake rest may not be as protective against interference as previously thought. Diekelmann, Büchel, Born, and Rasch (2011) compared awake versus sleep groups for the effect of cued reactivation on performance in an interference learning paradigm. Each group first learned object-location information, then underwent cued reactivation (through odor vs. vehicle), and finally completed a new location learning task (interference). The authors found that cued reactivation during sleep was associated with improved object-location memory after interference learning but, in the wake group, it was associated with reduced object-location memory because of interference learning. This finding, in line with other reconsolidation studies, suggests that awake reactivations may make memories more labile, allowing changes or updates to memory (Kuhl, Bainbridge, & Chun, 2012; Schiller et al., 2010; see Jardine, Huff, Wideman, McGraw, & Winters, 2022, for a review), but also creating opportunities for forgetting because of interference (Diekelmann et al., 2011). Another study from the same group later found that cued reactivation during awake rest benefits only the cued items, whereas benefits of cued reactivation is extended to uncued items from the same context during sleep (Oudiette et al., 2013), further supporting the notion that reactivation during wakeful and sleeping periods may have different effects on memory. Finally, a recent meta-analysis on targeted memory reactivation studies found that targeted-memory-reactivation-related memory benefits are more reliably observed during sleep than awake rest (Hu, Cheng, Chiu, & Paller, 2020). It is important to note that other dimensions of one's mental state, beyond just awake versus asleep periods, may have substantive impacts on the functions of reactivation. For instance, it has been suggested that reactivation effects may be more beneficial in an internally oriented compared with externally oriented brain states (Tambini, Berners-Lee, & Davachi, 2017). Others have shown that reactivation differs in low versus high reward conditions (e.g., Gruber et al., 2016), and in threatening versus safe contexts (de Voogd, Fernández, & Hermans, 2016). Together with these studies, our findings highlight the need for a more systematic investigation of the functions of reactivation as they arise in association with various mental states in humans.

The literature on memory reactivations in humans is very small, but surprisingly varied with respect to methodology. Laboratories conducting this type of research employ disparate tasks, stimuli, and memory testing methods (see Table 1), which potentially give rise to different results. In the spirit of transparency, we summarize features of our task that may have given rise to our results. First, using a grid for associative learning may have contributed to an increased level of interference, because of the relatively low discriminability of grid locations (compared with using close associates of the presented items). Most prior fMRI work on reactivation has focused on pairs of visual items to test associative memory (e.g., Tambini et al., 2010). Notably, the Diekelmann et al. (2011) study discussed above also utilized grid learning to specifically test how reactivations relate to later interference learning, and that study produced findings consistent with the present study with respect to wakeful reactivation, further supporting the idea that location-based learning may be particularly interference inducing.

Second, our stimuli were repeated several times during encoding. This may have altered the encoding patterns, or their strength, thereby leading to differences in their (co-) reactivation. However, we believe this variable is probably not essential to the pattern of our findings, given that previous work that has shown positive effects of reactivation on memory using varying numbers of repetition, from single shot learning (e.g., Staresina et al., 2013) to 30 repetitions (e.g., Deuker et al., 2013).

Third, our study tested recognition memory 3 times for all the learned image–location pairs. We acknowledge that this may have contributed to the current results in two different ways: On the one hand, as our behavioral findings suggest, there is a strong overall accuracy even after a 1-week delay, leaving us with limited behavioral variance across remembered and forgotten trials, which may bias the current results. On the other hand, it is possible that multiple testing sessions introduced additional interference over time, thereby leading to the forgetting that we observed over time.

Fourth, we did not find clear evidence of category-selectivity in our measures of reactivations in FFA (for face trials) and PPA (for scene trials). We suspect a unique property of our design, namely, the use of grid locations paired with each face and scene images, may have biased both regions to process the face and scene associates as objects (see Haist, Lee, & Stiles, 2010, for other examples of object processing in visual cortex). It is possible that this lack of category-specific reactivation in category-selective regions may relate to our unique forgetting effects: This lack of category specificity of cortical reactivations may point to a lack of reinstatement of important features of the events in the cortex, thereby biasing toward forgetting. Notably, others (e.g., Tambini et al., 2010) have failed to capture category-selectivity during reactivation as well, which points to a broader limitation in our understanding of how selective cortex engages in memory reinstatement.

Finally, our co-reactivation analysis identifies potential event replays on a TR-by-TR basis. We would like to acknowledge that the timing difference between the shorter duration of a TR (2 sec) and the original events (4 sec) may have contributed to some bias in the frequency of (co-)reactivations. Similar to previously employed TR-based analyses (e.g., Schapiro et al., 2018; Staresina et al., 2013), here, we rely on the assumption that event reactivations are temporally compressed, based on evidence from rodents (e.g., Davidson et al., 2009; Skaggs, McNaughton, Wilson, & Barnes, 1996) showing that memory reactivations during replay occurs on a faster scale than observed during encoding. Although recent studies have provided early evidence for temporal compression in human brain (e.g., Jeunehomme, Leroy, & D'Argembau, 2020; Schuck & Niv, 2019; Bonasia, Blommesteyn, & Moscovitch, 2016; Kurth-Nelson, Economides, Dolan, & Dayan, 2016), the time course of such temporal compression during memory replay in humans remains an open question. Future research is warranted to systematically test whether or how any of these factors could have contributed to these findings.

Current consolidation theories are, in general, critically lacking in the specification of boundary conditions: There is a lack of specificity regarding not only the types of (reactivation-like) processes that should occur, and how they should occur, but also in which conditions these processes are most likely to update, improve, or weaken memories. Table 1 summarizes the different features of the previous work showing memory outcomes related with event-specific spontaneous reactivations at awake rest. We believe some of these methodological differences highlighted in Table 1 reflect how vaguely defined theoretical concepts are translated into empirical choices. Thus, we conclude that future research should systematically test how alternative design approaches impact reactivation effects on memory and help define the boundary conditions for when and how reactivation, and co-reactivation, may improve or weaken event memories.

We would like to thank Zachary S. Heffernan for his contribution during data collection. We are also grateful to Dr. Charan Ranganath and Dr. Arielle Tambini for their early feedback on the methods.

Reprint requests should be sent to Büşra Tanrıverdi, Department of Psychology and Neuroscience Temple University 1701 N. 13th Street Philadelphia, PA 19122, or via e-mail: [email protected].

Data and analysis scripts are available upon request to the corresponding author. Additional information can be found here: https://osf.io/t3vug.

Büşra Tanrıverdi: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Software; Visualization; Writing—Original draft; Writing—Review & editing. Emily T. Cowan: Conceptualization; Formal analysis; Methodology; Software; Visualization; Writing—Review & editing. Athanasia Metoki: Conceptualization; Data curation; Investigation; Methodology; Project administration; Software. Katie R. Jobson: Data curation; Investigation; Project administration. Vishnu P. Murty: Conceptualization; Funding acquisition; Methodology; Supervision; Writing—Review & editing. Jason Chein: Conceptualization; Funding acquisition; Methodology; Resources; Supervision; Writing—Review & editing. Ingrid R. Olson: Conceptualization; Funding acquisition; Methodology; Resources; Supervision; Writing—Review & editing.

This work was supported by National Institute of Health grants to V. P. Murty (R01DA055259; R21DA043568), J. Chein (R01HD098097), and I. R. Olson (R01HD099165; R01MH091113; R21HD098509; and 2R56MH091113-11). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of Mental Health or the National Institutes of Health. In addition, this research includes calculations carried out on HPC resources supported by the National Science Foundation through major research instrumentation grant number 1625061 and by the US Army Research Laboratory under contract number W911NF-16-2-0189.

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance.