Of primary interest, we examined whether precise orientation information of the cued item is retained during WM retention in anatomically defined MTL regions of interest (ROIs; Figure 2A), including the entorhinal cortex (anterior-lateral, aLEC and posterior-medial, pMEC), the perirhinal cortex, para-hippocampus, and hippocampal DG/CA3, CA1, subiculum, as defined in the previous studies (Montchal et al., 2019; Reagh et al., 2017). Additionally, we chose the amygdala as a theoretically irrelevant but adjacent control region, because the involvement of the amygdala for emotionally neutral orientation information is expected to be minimal (Iwai et al., 1990). This allows us to gauge the observations in MTL ROIs while controlling for the signal-to-noise ratio in fMRI blood-oxygenation-level-dependent (BOLD) signals in deep brain structures.

As recent neural theories of WM have proposed that information retained in WM may not rely on sustained neural activation (Ester et al., 2015; Kamiński and Rutishauser, 2020; Rose et al., 2016), we inspected how the multivoxel activity pattern in each subject-specific ROI is correlated with the retained WM content predicted by the cued orientation gating (Figure 2B). We found that certain voxels in an ROI could respond more strongly to a particular cued orientation, even when the average BOLD activity across voxels does not show preferred coding for a certain orientation (see an example in Figure 2—figure supplement 1). We then assessed the consistency of these stimulus-related multivoxel activity patterns in the MTL and the amygdala control region based on stimulus-based representational similarity analysis. In this analysis, we correlated the angular similarity of every pair of cued orientation gratings with the similarity of the evoked BOLD patterns in these trials. The rationale is that if orientation information is retained within an ROI, the recorded neural data should track the relative angular distance between any two cued orientation gratings (hence fine discrimination Kriegeskorte and Wei, 2021). Informed by the previous research (Ester et al., 2015; Harrison and Tong, 2009), we performed this analysis using the raw fMRI BOLD signals from the middle 3TRs out of the 5-TR retention interval to minimize the contribution of sensory process or anticipated retrieval, hence maximizing the inclusion of neural correlates of WM retention (Postle et al., 2000).

In line with our prediction, we found that stimulus similarity for the cued item was significantly correlated with neural similarity across trials as compared with the null in both the aLEC (t(15) = 4.29, p=6.48e-04, p_Bonferroni = 0.0052, Cohen’s d=1.11, p_boostrap <0.001) and the hippocampal DG/CA3 (t(15) = 3.64, p=0.0024, p_Bonferroni = 0.019, Cohen’s d=0.94, p_boostrap <0.001; Figure 2C). In contrast, stimulus similarity for the uncued item across trials could not predict these neural similarity patterns in these regions as compared with the null (aLEC: t(15) = –0.20, p=0.85, Cohen’s d=–0.05; DG/CA3: t(15) = 0.06, p=0.95, Cohen’s d=0.02; p_boostrap’s>0.50). Furthermore, the evoked neural similarity patterns in these regions were significantly more correlated with the cued item as compared with the uncued item (aLEC: t(15) = 2.66, p=0.018, Cohen’s d=0.69, p_boostrap = 0.015; DG/CA3: t(15) = 3.64, p=0.0024, Cohen’s d=0.94, p_boostrap = 0.0016). While the rest of the MTL showed similar patterns, we did not obtain significant evidence in other MTL ROIs following the correction of multiple comparisons (see Supplementary file 1a for full statistics). Furthermore, neural evidence related to the cued item in the aLEC and DG/CA3 was significantly stronger than that in the amygdala control ROI. This was supported by a significant cue (cued vs. uncued) by region (combined aLEC-DG/CA3 vs. amygdala) interaction effect on the correlation between stimulus and neural similarity patterns (F(1, 15)=4.97, p=0.042, p_boostrap = 0.036). Together, these results suggest that delay-period activity patterns in the entorhinal-DG/CA3 pathway are associated with retrospectively selected task-relevant information, implying the presence of item-specific WM representation in these subregions.

To directly reveal the item-specific WM content, we next modeled the multivoxel patterns in subject-specific ROIs using an established inverted encoding modeling (IEM) method (Ester et al., 2015). This method assumes that the multivoxel pattern in each ROI can be considered as a weighted summation of a set of orientation information channels (Figure 3A). By using partial data to train the weights of the orientation information channels and applying these weights to an independent hold-out test set, one can reconstruct the assumed orientation information channels to infer item-specific information for the remembered item – operationalized as the resultant vector length of the reconstructed orientation information channel normalized at 0° reconstruction error (Figure 3—figure supplement 1). As this approach verifies the assumed information content based on observed neural data, its results can be efficiently computed and interpreted within the assumed model even when the underlying neuronal tuning properties are unknown (Ester et al., 2015; Sprague et al., 2018). This approach, therefore, complements the model-free similarity analysis by linking representational geometry embedded in the neural data with item-specific information under a model-based framework (Kriegeskorte and Wei, 2021; Xie et al., 2023b). On the basis of this method, previous research has revealed item-specific WM information in distributed neocortical areas, including the parietal, frontal, and occipital-temporal areas (Bettencourt and Xu, 2016; Ester et al., 2015; Rademaker et al., 2019; Sprague et al., 2016), which are similar to those revealed by other multivariate classification methods (e.g. support vector machine, SVM, Ester et al., 2015). We have also replicated these IEM effects in the current dataset (Figure 3—figure supplement 2).

Moving beyond these well-established observations in distributed neocortical structures, we found that the amount of reconstructed item-specific information for the cued item during WM retention was also significantly greater than chance level in two anatomically defined MTL subregions, aLEC (t(15) = 4.41, p=5.07e-04, p_bonferroni = 0.0041, Cohen’s d=1.14, p_boostrap <0.001) and the hippocampal DG/CA3 (t(15) = 4.73, p=2.68e-04, p_bonferroni = 0.0021, Cohen’s d=1.22, p_boostrap <0.001; Figure 3B). These effects were specific to the maintenance of the cued item, as information related to the uncued item was not statistically different from chance (aLEC: t(15) = –0.35, p=0.74, Cohen’s d=–0.09; DG/CA3: t(15) = 0.66, p=0.52, Cohen’s d=0.17; p_boostrap’s>0.50) and was significantly less than that for the cued item (aLEC: t(15) = 2.75, p=0.015, Cohen’s d=0.71, p_boostrap = 0.018; DG/CA3: t(15) = 3.83, p=0.0016, Cohen’s d=0.99, p_boostrap = 0.0023). Critically, the amount of information specific to the cued item in the aLEC and DG/CA3 was significantly greater than that in the amygdala control ROI, which is supported by a significant cue (cued vs. uncued) by region (combined aLEC-DG/CA3 vs. amygdala) interaction effect on IEM reconstruction outcomes (F(1, 15)=7.16, p=0.016, p_boostrap = 0.010).

Collectively, results from complementary analytical procedures suggest that the MTL’s entorhinal-DG/CA3 pathway retains precise item-specific WM content for a simple surface feature (e.g. orientation) to allow fine discrimination of different items in the feature space. As such, the stimulus-based prediction of neural similarity is highly correlated with the amount of reconstructed information based on IEM, even though these two analyses are based on different analytical assumptions (e.g. correlation between IEM and representational similarity analysis for the cued item, aLEC: r=0.87, p=0.000012, p_boostrap <0.001; DG/CA3: r=0.78, p=0.00037, p_boostrap <0.001; Figure 3—figure supplement 3).

Next, we examined the extent to which WM information retained in the MTL’s aLEC-DG/CA3 circuitry is related to an observer’s subsequent recall behavior. As the angular resolution of the reconstructed orientation information is 20° in the current study, our data therefore suggest that the MTL can distinguish similar orientation information in WM that is at least 20° apart. This neural separation should be consequential for later recall performance, in that trials with greater item-specific information reconstructed from the MTL should be associated with higher WM recall fidelity. To test this prediction, we grouped the trials from each participant into two categories. The first category contained small recall error trials, where participants made an effective recall response within one similar item away from the cued item (absolute recall error <20°; 149±3 trials [mean ± s.e.m.]). Another category contained larger recall error trials (27±3 trials) with absolute recall errors that were greater than 20° but smaller than the 3 standard deviations (SD) of the aggregated recall error distribution (Figure 4A). These trials would capture participants’ imprecise recall responses for the cued item, instead of those with an extra-large recall error that could be attributed to other factors such as attentional lapses (deBettencourt et al., 2019). The two identified categories of trials together account for about 98% of the total trials (i.e. 176 out of 180 trials).

We then performed the leave-one-block-out analysis to obtain trial-by-trial IEM reconstructions based on delay-period BOLD signals aggregated from the aLEC and DG/CA3. We averaged the IEM reconstructions from the small- and larger-error trials separately. Because trial counts between categories were not balanced, we resampled the data from the small-error trials based on the number of larger-error trials for 5000 times. We took the average of IEM reconstruction across iterations to obtain robust subject-level trial-average estimates with a balanced trial count across different behavioral trial types (Xie et al., 2020a; Yaffe et al., 2014). By contrasting these estimates at the subject level, we found that the small-error trials yielded significant IEM reconstructions for the cued item (t(15) = 4.50, p=4.21e-04, Cohen’s d=1.16, p_boostrap <0.001), whereas the larger-error trials did not (t(15) = 0.03, p=0.98, Cohen’s d=0.007, p_boostrap = 0.90; Figure 4B). Furthermore, the reconstructed WM information in the combined aLEC-DG/CA3 showed better quality in the small-error trials, as compared with that in the larger-error trials (t(15) = 2.45, p=0.027, Cohen’s d=0.61, p_boostrap = 0.032).

In addition to using an empirical criterion to separate in-memory trials from those extra-large error trials susceptible to occasional attentional lapses (deBettencourt et al., 2019), we have also tried another thresholding heuristic. As shown in Figure 1A, most trials from each participant fall within this 45° of absolute recall error (i.e. half of the 90° range), and the trials larger than this number are rare (~5 out of 180 trials). We, therefore, used 45° of absolute recall error as a cut-off to identify the imprecise recall trials that were greater than 20° but smaller than 45° of absolute recall error. We performed the same analysis to obtain trial-by-trial IEM reconstructions based on delay-period BOLD signals aggregated from the aLEC and DG/CA3 as outlined above, and then resampled the same number of trials to estimate the IEM reconstructions for the small-error and larger-error trials (<20° vs. 20° - 45° of absolute recall error). Consistent with the 3-SD heuristic, we found that the small-error trials identified by the 45° cut-off heuristic also yielded significant IEM reconstructions for the cued item (t(15) = 4.34, p=5.74e-04, Cohen’s d=1.12, p_boostrap <0.001), whereas the larger-error trials did not (t(15) = –0.69, p=0.50, Cohen’s d=–0.18, p_boostrap = 0.67). We then contrasted the difference in IEM reconstructions between these small- and large-error trials across participants. We found that IEM reconstruction for the cued item from the combined aLEC-DG/CA3 showed better quality in the small-error trials, as compared with that in the larger-error trials (t(15) = 3.41, p=0.004, Cohen’s d=0.88, p_boostrap = 0.008). Collectively, these results suggest that higher-quality WM representation in the entorhinal-DG/CA3 pathway during the delay period is associated with better subsequent recall fidelity and that this association is robust to the selection of cut-off scores for extra-large recall errors.

In sum, our data demonstrate that the MTL’s entorhinal-DG/CA3 pathway retains item-specific WM information, similar to that present in other distributed neocortical areas (Bettencourt and Xu, 2016; Ester et al., 2015). These results suggest that neural mechanisms underlying the fidelity of long-term episodic memory (Aimone et al., 2011; Bakker et al., 2008; Cappiello et al., 2016; Ekstrom and Yonelinas, 2020; Korkki et al., 2021; Marr, 1971; Reagh and Yassa, 2014; Yassa and Stark, 2011) are involved in representing precise item-specific WM content. Our data, therefore, provide broader insights into the fundamental constraints that govern the quality of our memory across timescales (Xie et al., 2023a; Xie et al., 2020b; Xie and Zhang, 2017a; Xie and Zhang, 2016).

Sixteen right-handed participants (mean ± s.e.m.: 21.32 ± 0.73 years old, 8 females) were recruited for the study with monetary compensation ($20/hour). This sample size was designed to be no smaller than that involved in the prior studies using similar experimental paradigms and analytical procedures (Bettencourt and Xu, 2016; Ester et al., 2015; Harrison and Tong, 2009). All participants reported normal or corrected-to-normal visual acuity and no history of neurological/psychiatric disorders or prior psychostimulant use. They provided written informed consent before the study, following the protocol approved by the Internal Review Broad of the University of California, Riverside (reference number: HS-17-035).

Participants performed an orientation visual working memory task adapted from previous studies (Ester et al., 2015; Harrison and Tong, 2009) inside an MRI scanner (Figure 1A). Briefly, on each trial, we sequentially presented two sine-wave gratings (~4.5° of visual angles in radius, contrast at 80%, spatial frequency at ~1 cycle per visual degree, randomized phase) at the center of the screen. Each grating appeared for 200ms, with a 400 ms blank screen in between. The two gratings had different orientations randomly drawn from nine predefined orientations (0–160° in 20° increments) and were >20° away from one another (see Figure 1—figure supplement 1). They were presented with a small random angular jitter (±1° to 5°). Following the offset of the second grating of each pair by 400ms, we presented a cue (‘1’ or ‘2’, corresponding to the first or second grating, respectively) for 550ms to indicate which grating orientation the participant should remember and maintain over an 8750 ms delay period. We instructed participants to remember only the cued grating and to ignore the uncued one. After the delay period, we presented a test grating initially aligned to a random orientation. Participants then pressed the response box buttons to continuously adjust the test grating until it matched the orientation of the cued grating based on their memory. We asked the participants to make a response within 3500ms following the onset of the test grating (averaged median response time across participants: 2929±156ms). After the response, we provided feedback to the participants by presenting a line marking the correct orientation, which was followed by an inter-trial interval of 3500 or 5250ms. Participants completed 10 blocks of 18 trials, yielding a total of 180 trials inside the scanner. Before scanning, they completed 2 blocks of 18 trials outside the scanner for practice. The cue position and the orientations of presented gratings were randomly intermixed within each block.

Under an effective set size of one item, participants’ recall performance was high (Figure 1B), with most recall errors centered around ±45° of the cued orientation (~97% of the trials) within the ±90° range. Hence, we retained all trials when investigating the amount of WM information in the recorded neural data during the delay period for multivariate analyses. We used the absolute recall error as a trial-level estimate of recall fidelity (Panichello and Buschman, 2021), assuming that large recall errors were driven by imprecise WM instead of other factors, such as occasional attentional lapses (deBettencourt et al., 2019; Xie and Zhang, 2017b). To minimize the contamination of these factors in linking the neural data with the behavioral data, we would focus on the trials where participants have recalled within the 3 SD of the aggregated recall error distribution (Figure 4A; see details in a subsequent section).

We acquired neuroimaging data using a 32-channel sensitivity encoding (SENSE) coil in a Siemens Prisma 3.0-Tesla scanner. We first acquired a high-resolution 3D magnetization-prepared rapid gradient echo (MP-RAGE) structural scan (0.80 mm isotropic voxels) and then functional MRI scans consisted of a T2*-weighted echo-planar imaging (EPI) sequence: TR = 1750ms, TE = 32ms, flip angle = 74°, 69 slices, 189 dynamics per run, 1.5×1.5 mm² in-plane resolution with 2 mm slice thickness, FOV read = 222 mm, FOV phase = 86.5%. This sequence was optimized for high-resolution functional MRI with whole-brain coverage for the scanner. Each functional run lasted 5 min and 30.75 s. At the end of the experiment, we acquired two additional scans with opposite phases to correct for EPI distortions (Irfanoglu et al., 2015).

We preprocessed neuroimaging data using the Analysis of Functional NeuroImages (AFNI) software (Cox, 1996). Briefly, functional data were de-spiked (3dDespiked), slice timing corrected (3dtshift), reverse-blip registered (blip), aligned to structural scan (align_epi_anat.py), motion-corrected (3dvolreg), and masked to exclude voxels outside the brain (3dautomask). To avoid introducing artificial autocorrelations in later analyses, functional data were not smoothed. For the same reason, we extracted the raw BOLD signals from the middle 3 TRs of the 5-TR retention interval for later analyses without fitting the data to the hemodynamic model (Ester et al., 2015). These raw BOLD signals were z-scored within each block/run, before extracting the TRs of interest. In particular, we convolved the data from the 5 TR delay period with a set of weights (i.e. 0, 1, 2, 1, 0) that resembled the TENT function in AFNI to maximize the inclusion of mid-delay activity for later analysis (Postle et al., 2000). This approach factors in 5–6 s of hemodynamic adjustment (Lewis-Peacock and Postle, 2008) and has been considered fundamentally conservative in estimating delay-period activity (Feredoes and Postle, 2007). This approach also provides a reasonable estimate for the BOLD response around a given TR with an improved signal-to-noise ratio without assuming the shape of the underlying hemodynamic response (Chen et al., 2015). We also performed the time-varying version of this analysis by shifting the peak of the TENT function over time (see Figure 3—figure supplement 5 for details).

To retain the consistency with the prior research, we defined participant-specific MTL ROIs (bilateral hippocampal DG/CA3, CA1, and subiculum, entorhinal/perirhinal cortex, and parahippocampus, see Figure 2A) based on the T1 image using the same segmentation algorithm from the previous studies (Montchal et al., 2019; Reagh et al., 2017). In brief, using the Advanced Normalization Tools (Avants et al., 2008), this algorithm aligned an in-house segmented template to each participant’s T1 image. This template contains manually labeled ROIs for hippocampal subfields (DG, CA3, CA1, subiculum) and other verified MTL subregions (aLEC, pMEC, perirhinal, and parahippocampus). The efforts to select and verify these MTL ROIs have been detailed in previous studies (Montchal et al., 2019; Reagh et al., 2018). In brief, in addition to the commonly identified perirhinal and parahippocampus ROIs, hippocampal subfields were manually identified and aggregated from a set of T1 and T2 atlas images based on prior harmonized efforts (Yushkevich et al., 2015). Entorhinal ROIs (aLEC and pMEC) were added to the template from a previous study (Maass et al., 2015). For functional analysis, we combined DG and CA3 subfields as a single label given the uncertainty in separating signals from them in fMRI data (Reagh et al., 2017). In addition, we also verified our findings in hippocampal subfields based on a different segmentation protocol via FreeSurfer (Iglesias et al., 2015), which yielded consistent findings (Figure 3—figure supplement 6). Therefore, our current observations are unlikely to be limited to a specific parcellation procedure of hippocampal subfields.

Furthermore, we identified subject-specific segmented amygdala as a control ROI based on participant-specific Freesurfer parcellation (Saygin et al., 2017). The amygdala is a part of the limbic system traditionally considered a central brain region processing emotion-laden information. Because the task stimuli (orientation gratings) and testing procedure (no reward manipulation) in the current study are emotionally neutral, the amygdala is therefore theoretically irrelevant for the current study (Iwai et al., 1990; Xie et al., 2022; Xie and Zhang, 2016). Furthermore, as its signal-to-noise ratio is similar to adjacent structures, the amygdala can serve as a control site for the observation in other MTL ROIs.

To examine whether MTL delay-period activity can distinguish different cued orientation gratings, we performed a stimulus-based representational similarity analysis (Kriegeskorte and Diedrichsen, 2019). The rationale is that if the recorded neural data contain information to allow fine discrimination of the cue item, the neural data should track the feature distance between any pair of cued items across trials to allow fine discrimination of these items (Kriegeskorte and Wei, 2021). Hence, we first calculated the stimulus similarity pattern across trials using 180 minus the absolute angular distance between the orientation labels of every two trials (Figure 2B, top panel). Next, we calculated the cosine similarity of the delay-period neural signals B across n voxels from the middle 3 TRs in every pair of trials (Figure 2B, bottom panel). This yields a trial-by-trial matrix in which the similarity between voxel response vectors Bi and Bj can be calculated as,S(Bi,Bj)=Bi ⋅ Bj||Bi|| ||Bj||

Finally, we correlated the neural similarity pattern and stimulus similarity pattern across trials (rank-order and Fisher’s transformed, Xie et al., 2018) to gauge how the recorded neural signals track the stimulus features across trials.

To decode item-level information from the raw BOLD signals (Ester et al., 2015), we first constructed a linear encoding model to represent orientation-selective responses in multi-voxels of activity from a given brain region. We did not impose any additional feature selection procedures other than using the anatomically defined ROIs to identify relevant multi-voxel features in this analysis (see Supplementary file 1c for the number of voxels/features included for each subject in each ROI). We assumed that the response of each voxel is a linear summation of 9 idealized information channels (Figure 2B), estimated by a set of half-wave rectified sinusoids centered at different orientations based on the tuning profile of orientation-sensitive neural populations. Hence, we formalized the observed raw BOLD signals B (m voxels ×n trials) as a weighted summation of channel responses C (k channels ×n trials), based on the weight matrix, W (m voxels ×k channels), plus residual noise (N),B= WC+N

Given B₁ and C₁ from a set of training data, the weight matrix can be calculated as,W= B1C1TC1C1T-1

The training weight matrix W was used to calculate a set of optimal orientation filters V, to capture the underlying channel responses while accounting for correlated variability between voxels (i.e. the noise covariance), as follows,Vi= ∑i−1WiWiT∑i−1Wi

where Σi-1 is the regularized noise covariance matrix for channel i (1–9), estimated as,Σi−1= 1n1−1εiεiTεi= B1-WiC1, i

Here, n₁ is the number of training trials, and εi is a matrix residual based on the training set B₁ and is obtained by regularization-based shrinkage using an analytically determined shrinkage parameter. Next, for the independent hold-out test dataset B₂, trial-by-trial channel responses C₂ are calculated as follows,C2= VTB2

We used a leave-one-block-out cross-validation routine to obtain reliable estimate channel responses for all trials. For each participant, in every iteration, we treated all but one block as B₁ and the remaining block as B₂ for the estimation of C₂. This analysis yielded estimated channel responses C₂ for each trial, which were interpolated to 180° and circularly shifted to a common center (0°, by convention). We reconstructed these normalized channel responses separately using orientation labels of the cued item, the uncued item, and shuffled orientations. We then quantified the amount of item-related information (R) by converting the average channel response (z) to polar form given ψ as the vector of angles at which the channels peak (z=Ce2iψ). We then projected them onto a vector with an angle of 0°,R=|z|cos⁡(arg(z))

With whole-brain coverage, we performed an additional searchlight procedure in combination with the IEM analysis to replicate the previous findings (Ester et al., 2015). First, we normalized participants’ brain data to an MNI template using the Advanced Normalization Tools. Second, we defined a spherical ‘neighborhood’ (radius 8.0 mm) centered on voxels in a cortical mask containing only gray matter voxels. We discarded neighborhoods with fewer than 100 voxels. Last, we estimated item-related information (R) about the to-be-remember item based on the IEM analysis outlined above to assess WM information within each searchlight sphere. We obtained consistent findings as compared with the previous findings (Figure 3—figure supplement 2), suggesting the reliability of the current data.

To examine how the IEM reconstruction of the cued item in the aLEC-DG/CA3 pathway is associated with recall fidelity, we performed the IEM analysis based on data combined from the aLEC and DG/CA3 ROIs. Similar to the analytical framework outlined above, we split each participant’s data into random blocks of 18 trials and then perform a leave-one-block-out analysis to obtain IEM reconstructions for all trials in each block based on the weights trained from other blocks. As this analysis is agnostic to participants’ recall performance at this stage, if IEM reconstruction is not associated with participants’ recall fidelity, the reconstructed information channels should be comparable regardless of recall errors. To test against this prediction, we split participants’ data into small- and larger-error trials. First, as the angular resolution was at least 20° for any two items in the current design, we defined small-recall error trials as those in which participants had reported within one similar item away (absolute recall error <20°; 149±3 trials). Next, to separate larger-recall errors based on less precise WM representation from those attributable to attention lapses (deBettencourt et al., 2019), we adopted a widely-used thresholding heuristic to find potentially different categories of data points based on the empirical SD of a distribution. Specifically, in our current data, we first calculated the empirical SD (17.33°) of the aggregated raw recall error distribution from all subjects across 2880 trials (ranging from –90° to 90°), which captures the overall variability in participants’ recall performance without a priori model assumption. We then retained the larger-recall error trials within 20° to 3 SD of the recall error distribution (27±3 trials; Figure 4A). These larger-error trials presumably contain mostly imprecise recall responses, instead of infrequent extra-large errors that could be attributed to other factors like attentional lapses (deBettencourt et al., 2019). Considering that most of the trials have a recall error of ±45° out of the ±90° range in every subject by visual inspection (97% of the trials, Figure 1B), we have also used 45° of absolute recall error as a cut-off for extra-large error trials and obtained similar findings in subsequent analyses.

To balance the trial counts between these two categories of trials, we resampled the same number of trials based on the number of larger-error trials from the small-error trials for 5000 times. This resampling procedure has ensured that the average IEM reconstruction from the small-error trials was estimated based on the same number of trials as compared with the larger-error trials – an approach often used to obtain less biased estimates of neural measures across different behavioral trial types (Xie et al., 2020a; Yaffe et al., 2014). We contrasted the difference in IEM reconstructions for the cued item in the aLEC-DG/CA3 between these two categories of trials across participants.

We evaluated statistical significance based on conventional within-subject statistical procedures, such as paired-sample t-tests, with two-tailed p values. Similar results were obtained and verified based on non-parametric statistics (e.g. bootstrapped p values) that have few analytical assumptions (Good, 2013). In particular, we resampled participants’ data with replacement over 1000 iterations and calculated the empirical two-tailed p values (note: these p values can slightly vary across different iterations of resampling and those smaller than 0.001 are marked as p_bootstrap <0.001). We estimated the size of these effects based on Cohen’s d. Except for pre-defined contrast analysis (e.g. cued vs. uncued), we corrected for multiple comparisons by using Bonferroni correction with an alpha level set as 0.05 (Rosenthal and Rubin, 1983). For visualization of variability in mean estimates, we have used the standard error of the mean across participants (s.e.m.), namely the standard deviation of a measure divided by the square root of sample size, as error bars (or areas) in Figures 2—4.