The human hippocampus can pattern separate memories by meaning Authors: Alex Ilyés, Borbála Mónika Brosig, György Mező, Attila Keresztes Published: 2026 ABSTRACT SignificanceMemory research traditionally distinguishes semantic memory, which supports knowledge about the world, from episodic memory, which supports richly contextualized recollection of events. As Tulving originally emphasized, this distinction was not intended to imply fully independent systems, and accumulating evidence suggests shared neural operations across both. Using state-of-the-art high-resolution MRI combined with the manipulation of semantic similarity quantified using a distributional language model, we found that hippocampal pattern separation, which supports mnemonic discrimination of episodic memories by orthogonalizing similar neural inputs into nonoverlapping representations, can also resolve overlaps in semantic content. These findings provide direct evidence for a neural operation bridging the two memory systems. Hippocampal pattern separation is a neural operation responsible for creating orthogonalized representations from overlapping inputs (1–3). Converging evidence from animal (4–6), human (7–9), as well as clinical (10–12) studies suggest that the dentate gyrus (DG) subfield of the hippocampus is critical in establishing this sparsification (13–15). The resulting sparse signal is then forwarded via mossy fibers to the cornu ammonis 3 (CA3) (16) where separated representations are also observable (17, 18). At the cognitive level, in humans, pattern separation supports mnemonic discrimination (19–21)—sometimes referred to as “behavioral pattern separation” (22, 23)—i.e., successfully detecting novelty of lures that are highly similar to already encoded memory traces. This is often assessed via the mnemonic similarity task (MST) (19, 22), which tests mnemonic discrimination of highly similar pictures of everyday objects. At the neural level, the most typical operationalization of pattern separation in humans leverages the alteration of region-specific blood oxygen related (BOLD) signal upon repetition of a stimulus [i.e., repetition suppression (RS) effect] (2, 24, 25). This “repetition sensitivity approach” (26) allows identification of regions [within and beyond the HC (27)] that treat a stimulus as a repetition of a previously presented stimulus. Importantly, it further allows to distinguish clusters within these repetition sensitive regions that respond to similar lures as repetitions from those that respond to similar lures as novel stimuli. A core assumption of this approach is that the regions treating similar lures as novel perform pattern separation. Parametrically manipulating the degree of similarity between lures and encoded traces allows for a fine-grained measurement of such differences in RS (8). Based on converging evidence from human fMRI studies, regions that perform pattern separation, should exhibit 1) attenuated RS to similar lures, suggesting that these are treated as novel rather than repeated items (7–9, 28), 2) an activation pattern with small differences in input leading to large differences in output with moderate or no further effect of increased input differences (8, 26, 29–31), and 3) sensitivity to mnemonic discrimination success, such that falsely recognized lures elicit decreased activity compared to correctly rejected ones (19, 32–35). Studies using the MST have provided evidence for hippocampal pattern separation in the DG (7, 9). However, due to the complex nature of everyday objects, confounding factors of perceptual, conceptual, and spatial similarity make it hard to infer which representational dimensions are being separated by the hippocampus. Despite proposals that the DG may have a “universal pattern separator” (p. 299) role across domains (20), evidence remains scarce. Azab et al. (20) used a design inspired by the MST using spatially organized sequences of stimuli and found attenuated RS for similar stimuli in the DG consistent with neural pattern separation independent of temporal or spatial manipulation. Based on the framework of Cowell et al. (36), we believe that further human studies implementing the standard procedure of the MST with varying representational content are important to precisely understand what representations hippocampal pattern separation can operate on. The hippocampus is often designated as a “higher-level layer in the ventral visual stream” (p. 905) (37), which aligns well with the use of visual information for interference resolution. In fact, studies found a direct relationship between visual discrimination and mnemonic discrimination in humans (38, 39). However, recent advances in modeling the semantic system have uncovered the role of HC in accessing and manipulating semantic information (40). For instance, hippocampal patients list fewer semantic features for basic concepts compared to healthy controls (41), and fail to navigate toward distantly related associations (42). The hippocampus is also implicated in semantic dementia (43) with distinctly left lateralized disruptions of functional connectivity compared to bilateral disruptions in Alzheimer’s disease (44). Healthy populations also show a reliance on HC during navigation of conceptual spaces (45–48). Additionally, the HC is sensitive to both conceptual and perceptual similarity of everyday objects, even when modeling these factors separately (49–52). In light of these findings, a key open question is whether the orthogonalization of information based on meaning derived from verbal information also recruits hippocampal pattern separation. Two earlier studies investigated mnemonic discrimination of verbal stimuli: Ly et al. (53) probed mnemonic discrimination of verbally presented, semantically similar objects (single nouns, e.g., “rattle” vs. “bell”), while Ilyés et al. (54) used the “semantic MST” (sMST) also probing mnemonic discrimination of verbally presented concepts (e.g., “zoo”). Crucially, in the latter study, similarity was parametrically manipulated in finer conceptual details (e.g., “exotic zoo” vs. “strange zoo”) and not broad concept categories (e.g., apple vs. pear), analogous to how the original MST manipulates perceptual details of the same object image. Performance on the sMST also correlated with standard MST scores in the same sample [r = 0.46, P < 0.001 from Ilyés et al. (54)], raising the possibility of a shared pattern separation mechanism. Here, we used the sMST (54) to test whether hippocampal pattern separation supports mnemonic discrimination of adjective-noun phrases. We followed the review-based guidance of Liu et al. (55) to ensure that our test is a valid measure of pattern separation for semantically similar episodic memory representations. First, we showed repeated and modified items during both the incidental encoding and the recognition phase of the task to reduce the neural confound of a “recall to reject” strategy influencing RS measures (56). Second, we parametrically manipulated semantic similarity via cosine similarity derived from a word2vec word-embedding (57, 58). Finally, we controlled for covariates that were either relevant for general cognition (processing speed), semantic cognition (vocabulary) or were sensitive to mnemonic discrimination of semantically similar phrases in our earlier study (phonological awareness). We used high-resolution functional MRI (fMRI) to detect pattern separation-consistent variations of the BOLD signal in the HC during encoding and recognition of adjective-noun phrases parametrically manipulated in semantic similarity. We expected that 1) repetition sensitive regions in the HC, specifically the DG-CA3, will show attenuated RS for lures, 2) even small differences in meaning will lead to differences in encoding related BOLD signal in the DG-CA3 with larger differences in meaning leading to little or no further difference in BOLD, 3) encoding related activity in repetition sensitive regions will be associated with behavioral mnemonic discrimination performance, and 4) repetition sensitive regions will differentially respond to correctly rejected versus falsely recognized lures. Results Mnemonic Discrimination Performance Decreases as a Function of Semantic Similarity. First, using the sMST (Fig. 1A) we tested whether semantic similarity affects mnemonic discrimination (Fig. 1 B and C) of lures during recognition. A one-way repeated measures analysis of covariance (ANCOVA) on d-primes (by type: target vs. close lure, target vs. distant lure, target vs. foil) with vocabulary, processing speed, and phonological awareness as covariates, suggested robust differences (F(2, 46) = 174.6, P < 0.001, η2 = 0.58) in discrimination performance for lures at different levels of semantic similarity and foils from targets (Fig. 1D). Crucially, all d-primes were significantly different from one another according to Bonferroni-corrected paired t tests (target-close vs. target-foil: t(27) = −17.9, P < 0.001, CI [95%] = [−1.34 to −1.07]; target-distant vs. target-foil: t(27) = −11.7, P < 0.001, CI [95%] = [−0.99 to −0.69]; target-close vs. target-distant: t(27) = −5.24, P < 0.001, CI [95%] = [−0.51 to −0.22]) reflecting that mnemonic discrimination was lowest for close lures, followed by distant lures, and highest for foils. Task design and procedure of the sMST and mnemonic discrimination performance. (A) First, participants incidentally encoded adjective-noun phrases (first presentation, e.g., exotic zoo, n = 175), with nouns repeated during encoding either paired with the same adjective (exact repeat, n = 50) or paired with a semantically similar adjective (modified repeat, n = 50) that was either close (close modified, e.g., strange zoo, n = 25) or distant (distant modified, e.g., alternative zoo, n = 25) in meaning to the previously seen pair (for more details see Methods, Behavioral methods, Materials). In a subsequent recognition phase, participants made old/new judgments on repeats of previously seen first presentations paired with the same adjective (targets, n = 50), repeats of previously seen nouns paired with semantically similar adjectives that were not seen during encoding (lures, n = 50), and phrases with both novel adjectives and nouns (foils, n = 50). The distribution of responses for each recognition condition is plotted at the bottom excluding trials with no responses. The x axis shows the conditions during the recognition phase, and the y axis shows the percentage of either “old” responses (in red) or “new” responses (in blue). (B) We parametrically manipulated the semantic similarity of modified adjective-noun phrases sharing the same noun via calculating the cosine similarity (Stimuli) between vectors in a Hungarian word2vec word-embedding (58). (C) Histograms of cosine similarity to the first presentation of each semantically close (in blue) and distant modified phrases (in yellow) separately (note that both encoding modified repeats and recognition lures were selected from this set). We introduced a cut-off at 0.65 during the stimuli creation. (D) Boxes represent two-response LDI, calculated by subtracting the probability of responding old for a given condition from the probability of responding old to targets (d′ or d-prime). Lower d′ means that it is harder to discriminate items in specific conditions. (E) Slopes of semantic similarity predicting percentage of old response (false alarm) for lures (collapsed and binned as close and distant lures) during encoding. Note that cosine similarity was not in interaction either with lure type or with encoding repeat type. Next, we assessed the effect of fine-grained semantic similarity as a continuous variable on mnemonic discrimination. Using a “probit” generalized linear mixed effect model (GLMEM), we found that the probability of false alarms for both close and distant lures increased as a function of semantic similarity (estimate = 5.13, SE = 1.81, P < 0.01, CI [95%] = [1.59 – 8.67], Fig. 1E). We report detailed methods for these analyses in Behavioral Methods and supporting results in SI Appendix, Supplementary Results 1. RS Sensitive Clusters in the Hippocampal Head Differentially Activate for Exact Repeats and Modified Repeats During Encoding. To assess neural signatures of pattern separation at encoding, we tested for the specificity of RS to exact repeats compared to modified repeats during encoding using a three-way repeated measures ANCOVA on parameter estimates extracted from two repetition sensitive clusters (described in detail in Analysis) of the hippocampal head bilaterally (Fig. 2A.) localized predominantly within the DG-CA3 (SI Appendix, Table S2). Importantly, the anterior location of the uncovered clusters aligns with findings in previous pattern separation studies (Fig. 2C, for more details see SI Appendix, Supplementary Methods). We found main effects of condition (F(1, 25) = 16.27, P < 0.001, η2 = 0.05), and order (F(1, 25) = 50.95, P < 0.001, η2 = 0.14), as well as the critical two-way interaction between order and condition (F(1, 25) = 12.81, P = 0.001, η2 = 0.03) indicating significant RS that differed across exact and modified repeats. We also found an order and cluster (F(1, 25) = 5.68, P = 0.02, η2 = 0.006) interaction, showing that the RS effect differed between the clusters. Post hoc paired t tests revealed that first vs. second presentation of modified phrases differed significantly (t(55) = 3.44, P = 0.002, CI [95%] = [0.03 to 0.11]) across clusters, although the RS effect was attenuated (Fig. 2B) compared to exact repeats (t(55) = 8.86, P < 0.001, CI [95%] = [0.15 to 0.23]). Consistent with hippocampal pattern separation, exact repeats and modified repeats also differed significantly (t(55) = −5.45, P < 0.001, CI [95%] = [−0.18 to −0.08]), the latter evoking a higher signal. We report detailed methods for this analysis in Analysis and all covariate related effects in SI Appendix, Supplementary Results 1. Attenuated RS across repetition sensitive clusters in the hippocampal head consistent with previous findings of pattern separation. (A) We identified two repetition sensitive clusters at a liberal group level threshold (Analysis) showing the RS effect (first presentations of exact repeats–exact repeats) bilaterally in the hippocampal head (in white): left hippocampal head (x = −21, y = −16, z = −19, peak z-score = 3.02, 172 voxels), right hippocampal head (x = 21, y = −13, z = −18, peak z-score = 2.962, 230 voxels). Note, that the majority of voxels (110 in the left hemisphere, 148 in the right hemisphere) were localized in the medial part of the DG-CA3. Automatically segmented hippocampal subfields in the hippocampal head were transformed in a study-specific template space overlaid on the template T1 (Hippocampal Subfield Segmentation) in MNI space (59) as a background (blue—DGCA3, red—CA12, green—subiculum). (B) Facets show mean percent signal change values in the regions of interest (ROI) during incidental encoding. Error bars denote within-subject SE. A lack of three-way interaction indicates that RS for modified repeats is not reliably different in the ROIs. (C) Circles (current study in red) represent cluster peaks showing pattern separation-consistent activity in MNI space. The “x” symbols show peak activation in studies where y coordinates were defined in Talairach-space, whereas x and z coordinates were selected based on visual inspection of figures (for more details see SI Appendix, Supplementary Methods). For visualization, we used nibabel (60) (version 3.9.4) and matplotlib (61) (version 5.3.3) to project the hippocampus mask at 50% probability from the Harvard-Oxford atlas (62–65) in MNI space (59). Neural Response Is Increased for Semantically Close Modified Phrases Compared to Exact Repetitions, with No Further Difference between Semantic Similarity Bins. Next, to test whether even small input differences lead to large output differences with further input differences having little or no further effect, we modeled percent signal changes at encoding across the three repetition conditions (exact, close modified and distant modified repeats). Phrases were binned into close and distant modified repeats based on whether the same noun was modified with a semantically close or distant related adjective during encoding. We ran a two-way ANOVA predicting percent signal change by repetition type (exact, close and distant bins) in interaction with clusters and found a significant effect of repetition type (F(2, 56) = 7.35, P = 0.001, η2 = 0.06), showing differences between exact and semantically close (t(57) = −3.57, P < 0.001, CI [95%] = [−0.15 to −0.04]), as well as exact and semantically distant (t(57) = −4.08, P < 0.001, CI [95%] = [0.17 to 0.06]), but no significant difference between semantically close and semantically distant repeats (t(57) = −0.56, P = 0.58, CI [95%] = [−0.07 to 0.04]) (Fig. 3). When probing anatomically defined regions of interests, this pattern was similar in the CA1-2 and DG-CA3 in the right hippocampal head, but the exact repeat vs. semantically close modified repeat contrast reached significance in the CA1-2 (t(28) = −3.48, P = 0.002, CI [95%] = [−0.1 to −0.02]) only, and not in the DG-CA3 (t(28) = −1.82, P = 0.08, CI [95%] = [−0.1 to 0.006]; for details see SI Appendix, Supplementary Results 2). Attenuation of RS for exact repeats is evident for semantically close as well as semantically distant modified repeats, with no difference between close and distant modified repeats. Facets show mean percent signal changes in the ROIs during incidental encoding for each repeat condition (exact, close, distant). Error bars denote within-subject SE. Neural Signal of Pattern Separation in the Right Hippocampal Head During Encoding Is Associated with Behavioral Mnemonic Discrimination Performance During Recognition. Next, to test the association between neural pattern separation and behavioral mnemonic discrimination, we correlated the participant-level differences between percent signal changes of modified and exact repeats with d’ values for lure vs. target discrimination per participant. The neural index of pattern separation significantly and positively correlated with mnemonic discrimination performance specifically in the right hemisphere cluster (r = 0.44, P = 0.034, CI [95%] = 0.09 to 0.69), and positively but not significantly in the left hemisphere cluster (r = 0.29, P = 0.13, CI [95%] = −0.09 to 0.59; Fig. 4). Importantly, the correlations did not differ statistically [t = −1.01, P = 0.32, based on Steiger (66)], suggesting that this association may not be specific to the functional cluster in the right hemisphere. Neural index of pattern separation at encoding positively correlates with mnemonic discrimination of lures at recognition in the right hippocampal head. Facets show correlations between modified and exact repeat difference (neural index of pattern separation) and target vs. lure d′ (behavioral index for mnemonic discrimination) in repetition sensitive clusters. The red line shows significant correlation at Bonferroni-corrected P < 0.05. Note that the correlations do not statistically differ (t = −1.01, P = 0.32) from each other. Clusters Sensitive to Repetition at Encoding Do Not Track Mnemonic Discrimination Success within Lures but Differentiate Correctly Rejected Lures from Correctly Rejected Foils During Recognition. Finally, to establish a link between encoding-related and recognition-related neural signals of pattern separation, we tested whether clusters in the HC showing pattern separation at encoding differentiate lure trials during recognition as a function of discrimination success, comparing lure correct rejections versus lure false alarms (32–35). A key difference to prior work was that we used functional masks derived from a separate encoding phase to assess recognition-related activation. Because participants made few false alarms on average, we excluded those who did not produce at least five false alarms, and we ran Bonferroni-corrected paired t tests on the difference between lure correct rejections and lure false alarms. We found no significant differences in either of the clusters (ps > 0.15, reported in detail in SI Appendix, Supplementary Results 8). As a proof-of-concept analysis, to test whether we can detect any BOLD signal modulation by experimental conditions during recognition (i.e., target, lure, foil) in clusters uncovered during encoding, we ran a repeated measures two-way ANOVA on percent signal changes with fixed effects of condition and cluster. We found a main effect of condition (F(2, 58) = 6.97, P = 0.002, η2 = 0.04) driven by a difference between foil and lure correct rejections across clusters (t(59) = 4.61, P < 0.001, CI [95%] = [0.07 to 0.173]), although lure correct rejections were also lower than target hits (t(59) = −2.03, uncorrected P = 0.047, CI [95%] = [−0.12 to −0.001]). However, this comparison did not survive Bonferroni-correction. To determine whether this difference was due to a signal enhancement for novel items or a suppression for lures, we ran Bonferroni-corrected t-tests and found that activation for lures were significantly below implicit baseline (t(59) = −2.77, P = 0.015, CI [95%] = −0.157 to −0.025), while foils did not show a difference (P = 0.67), suggesting a signal suppression for lures (Fig. 5). Finally, as pointed out by one of the reviewers, the orthogonal manipulation of encoding (exact repeat vs. modified repeat) and recognition conditions (target vs. lure) allowed us to analyze their separate contributions to recognition signals. We report results of these admittedly exploratory analyses, including caveats, in SI Appendix, Supplementary Results 7. Repetition-sensitive clusters differentiate correctly rejected lures and foils during recognition. Facets show mean percent signal changes of correctly rejected foils, lures, and correctly recognized targets during recognition in both repetition sensitive clusters detected during encoding. Error bars indicate within-subject SE. Discussion We investigated how hippocampal pattern separation supports mnemonic discrimination of verbally presented memory traces that were parametrically manipulated in the similarity of their meaning via a word-embedding. We specifically aimed to uncover established indices of mnemonic discrimination at the behavioral level and pattern separation at the neural level, while we presented novel, repeated, and semantically modified stimuli during both an incidental encoding phase and a recognition phase separately. As an overview, we uncovered the following results with respect to our predictions: 1) RS was attenuated during the encoding phase for semantically modified phrases compared to exact repeats in repetition sensitive clusters in the bilateral hippocampal head (localized predominantly within the DG-CA3), 2) the attenuated RS appeared already at high semantic similarity with no additional difference compared to low similarity phrases, 3) a proxy of neural pattern separation—the difference in percent signal change between semantically modified and exact repeats—at encoding was positively associated with mnemonic discrimination performance at recognition, and 4) activation during recognition was independent of the discrimination success of lures modeled in repetition sensitive clusters found at encoding. By employing a “repetition sensitivity” approach (26) we showed that clusters in the hippocampal head show RS when encountering exactly repeated adjective-noun phrases and show an attenuated RS when the noun is repeated with a semantically similar adjective (e.g., strange zoo relative to exotic zoo). The uncovered pattern—i.e., lower activation for exact repeats compared to both first presentations and modified repeats—is considered to be a neural pattern separation proxy in humans (7–9, 28) and was recently also found in rhesus monkeys (67), whereby a similar firing pattern of single neurons for novel and lure items different from that of repeats predicted mnemonic discrimination success. The neural index we used—i.e., the difference of signal change for repeated and modified trials at encoding—was also associated with behavioral mnemonic discrimination performance at recognition. We consider this association as evidence that the RS-based pattern separation proxy we found is indeed a pattern separation signal that helps later mnemonic discrimination. Our results extend existing evidence by demonstrating pattern separation-consistent signals in the hippocampus when mnemonically discriminating highly similar verbal material instead of pictures of everyday objects. Following the framework of Cowell et al. (36), we believe that pattern separation as a neural operation takes part in the discrimination of highly similar traces regardless of their representational content. This is in line with other studies showing that the HC or adjacent MTL regions employ pattern separation regardless of content, such as rotation (29), temporal sequence (20), emotional information (32), novel abstract shapes in a DG-lesioned case study (68), objects and their contexts independently (69), and nonmeaningful fractals and their spatial location within the same context (39). Our study also closely resembles the original perceptual MST, but instead of showing a visually slightly modified picture of the same object as lures, we kept the object constant by repeating the noun modified by another semantically relevant adjective. Thus, the difference between neural signals for exact and modified repeats reflects sensitivity to fine-grained meaning information (70), and not a difference in semantic category or gist. The involvement of the anterior HC in pattern separation based on fine-grained information is seemingly in contrast with prevalent models suggesting specialization for gist-like processing in the anterior and for fine-grained processing in the posterior HC (71, 72). However, a recent meta-analysis (73) found that the anterior HC is specifically more active during encoding compared to the posterior HC, which may explain the anterior involvement in pattern separation during an encoding phase. In fact, multiple studies have detected anterior hippocampal clusters (sometimes exclusively so) involved in pattern separation assessed specifically as RS-related attenuation for similar items during encoding (7, 9, 74–79). Anterior HC involvement in pattern separation involving preexisting semantic representations was further supported by a recent study showing a learning-related increase in representational dissimilarity between subsequently remembered novel-famous face pairs within the anterior HC (80). Along similar lines, representing an object via meaning compared to a picture involves representation of amodal semantic information. Thus, in line with earlier findings of anterior HC involvement in semantic impairments (81) and memory-based neural pattern differentiation during verbal cueing (82), pattern separation of such amodal semantic information may have facilitated anterior HC involvement. We also detected an RS attenuation already for small differences in meaning and observed no significant further difference in attenuation for large differences in meaning. This is consistent with pattern separation-related activation patterns reported elsewhere, including findings showing no difference between lure similarity bins in young adults during encoding (8, 31, 76, 83) or retrieval (26), and studies fitting a power function to a sharp activation increase for high similarity levels relative to exact repeats, with no further differences across increasing similarity levels (27, 29). Although our repetition-sensitive clusters were predominantly located in the anterior DG-CA3, subfield-specific region of interest analyses suggested that the specific pattern of RS attenuation for semantically close modified repeats was significant in the right anterior CA1-2 only. Within the right anterior DG-CA3 only the low similarity bin differed significantly from exact repeats, whereas in the left hemisphere no comparisons reached significance. The inconsistency of anatomical results may partly be due to challenges of delineating anterior hippocampal subfields, especially in the medial uncal region, where the measurement may reflect a mix of signals from different subfields (84). We also did not find a significant difference between semantically close and distant lures at recognition (SI Appendix, Fig. S5), whereas behavioral results revealed a difference in discrimination of similar lures as a function of semantic similarity, replicating earlier results (54). Importantly, univariate contrasts of a priori determined similarity bins may simply not be sensitive enough to detect fine-grained effects of semantic similarity—measured as a continuous variable—on neural signals of pattern separation. Unfortunately, this was not a viable analysis option because of a design choice: Participants saw different counterbalanced stimuli lists to ensure that effects were independent of particular nouns and their semantic distances. This prevented us from being able to reliably assess group-level effects of a continuous similarity variable on neural estimates at either encoding or retrieval (50, 52, 85) in association with variations in memory performance (42, 86). Alternatively, the isolation of semantic content could have promoted extrahippocampal structures to process semantic information in parallel with perceptual processing (87–89) changing the representation that the HC operated on. On a related note, whether an exact or a modified repeat was presented during encoding did not affect discrimination of lures during recognition. This is surprising, because reencoding of exactly repeated stimuli compared to ones with modified repeats should lead to a bias toward recall rather than familiarity (90). However there might be a strong familiarity-based decision making behind the completion of the task regardless of repetition type [as found in the perceptual MST (91)], which might have also occluded parametric differences in hippocampal neural signals across similarity bins during recognition. In our encoding-derived functional mask, we did not detect an alternative signal of pattern separation, i.e., a suppressed activation for lure false alarms compared to correct rejections during recognition (19, 32–35). This signal is often attributed to the successful orthogonalization of information to correctly reject similar lures. One main reason for the lack of effect could have been due to design limitations. Since we aimed to assess neural signals of pattern separation both during encoding and recognition, for the sake of efficiency, we repeated and semantically altered the same stimuli set across the two phases. Thus, univariate estimation of lure correct rejections and false alarms contained items that were repeated twice or were modified during encoding. Reagh et al. (92) showed that repeating items three times versus showing them only once during encoding specifically decreased the correct rejection vs. false alarm difference for lures during recognition. Another limitation was the use of a functional cluster derived during the encoding phase to test signal alterations during the recognition phase. The sensitivity of this analysis may be dampened by the fact that task context can change the functional relevance of voxels within the HC (93). Our follow-up exploratory analyses nonetheless uncovered experiment-related alteration in the encoding-based functional clusters during recognition, specifically lower activation for correctly rejected lures compared to correctly rejected foils, suggesting that these clusters did have functional importance in both phases. The uncovered difference may be considered as a “novelty signal” (34) for foils compared to lures as foils were novel concepts defined by unseen nouns during encoding. Alternatively, because the percent signal change for lure correct rejections was lower than the baseline, it could be that we detected a left-over RS for the repetition of the conceptual gist information, in line with no difference between lure correct rejections and false alarms. In fact, Motley and Kirwan (29) found similar results when assessing RS-related signal alteration in an MST while manipulating encoding versus recognition instructions. In the left HC, they identified a cluster showing pattern separation-consistent increase following a power function across lure similarity levels during encoding, but an inverse power function during recognition, such that activity was reduced for the most similar lures relative to first presentations and least similar lures. They interpreted this inverted function as an RS-driven “pattern completion” signal during recognition. If we consider our foils as the least similar phrases to already encoded phrases, then suppressed signal for lures relative to foils during recognition could also be interpreted as a pattern completion signal. Indeed, an RS effect may persist between encoding and retrieval phases: Neural RS (94, 95) as well as behavioral priming (96) often survive multiple days of delay. Alternatively, new information may have been assimilated into cortically stored representations in parallel with hippocampal pattern separation. This is consistent with a study by Bein et al. (80) showing that successful subsequent recognition of famous-novel face pairs was associated with reduced similarity between item-specific multivariate patterns within the HC (signaling pattern separation), but increased similarity between item-specific multivariate patterns within the cortex. In another study of the same group, using the same design, novel faces were also harder to retrieve when paired with famous ones with preexisting representations (97). Together, these studies suggest that assimilated cortical representations could have contributed to both the decrease of HC activation during retrieval and the increase in false alarms with increasing semantic similarity observed in our study. The multiple potential mechanisms listed above highlight the challenges of interpreting retrieval-related signal differences. Task context can change not only how neuronal ensembles are engaged but also the neural operation driving the RS effect, as well as its behavioral relevance. These challenges call for studies specifically designed to separate the assessment of encoding vs. retrieval-related signals of pattern separation. In principle, our design allowed us to assess recognition signal alterations within the HC as a function of both encoding and recognition conditions to distinguish competing views of pattern separation: 1) an on-line pattern separation account predicting pattern separation both during encoding and retrieval due to no encoding or storage capabilities of the DG (98), 2) an encoding-specific pattern separation account (99, 100) predicting that pattern separated representations reduce the need for DG engagement during retrieval, consistent with rodent evidence showing encoding deficits but relatively preserved consolidation and retrieval in DG lesioned rats (101, 102). Post hoc exploratory analyses (SI Appendix, Supplementary Results 7) provide tentative support for model (2) by showing that correctly rejecting lures of strongly encoded representations and recognizing targets that were already pattern separated at encoding produce lower signals than strongly encoded targets, hinting at disengagement of the pattern separation circuit. Because our study was not designed as a test between these models, we encourage the use of similar cross-over designs probing other representational modalities to provide further evidence for dissociating pattern separation during encoding and recognition. Importantly, the pattern separation-consistent clusters in our study were predominantly located in the DG-CA3, and considerably fewer voxels overlapped with the subiculum and CA1-2, which is in-line with a DG-specific pattern separation account. However, this localization may be imprecise due to the use of a group-level template, inaccuracies in spatial normalization, and the lack of a validated segmentation protocol for the hippocampal head. In fact, the critical pattern separation proxy did not reach significance when using anatomical masks as ROIs (SI Appendix, Supplementary Results 3). Critically, the effect did not appear in the hippocampal body for which a recently published consensus-based, harmonized, valid and reliable protocol allows us to separately delineate CA3 and DG (103). There have been similar results reporting inconsistent pattern separation signals in anatomical, but not functional ROIs (9) and pattern separation-consistent signals have also been observed in the subiculum (35). This highlights further need toward harmonized and validated protocols of hippocampal subfield segmentations in the hippocampal head (104), and also suggests that even hippocampal subfields might be functionally heterogeneous, with lower-level neuronal ensembles taking part in different operations [similar evidence available from cell type–based functional heterogeneity in mice (5, 105)]. Our results are promising in uncovering the role of hippocampal pattern separation in remembering fine-grained semantic details, and point to several clear directions for future work. First, separate studies of incidental encoding and intentional recognition might allow for more items per bin to achieve higher statistical power for modeling effects separately. For instance, it would be interesting to assess the neural signal separately for exactly repeated phrases and repeats of only the broad conceptual information. Second, with temporal and spatial resolution used in our study, scanning at higher field strengths could allow for covering the full brain with the possibility of detecting extrahippocampal contributions to pattern separation based on semantic content. Finding a task-based link between hippocampal and anterior temporal function with the sMST—specifically the temporal pole, that has been shown to code semantic similarity (106)—might allow for integrating theories of gist-based false memories (107) and pattern separation failure (83), a link specifically interesting for studies of aging. Finally, multivariate analyses with richer parametric semantic manipulations (i.e., category cueing, construction of novel semantic spaces), could establish the specific scope of the HC in using remote, graded semantic information during episodic operations. Materials and Methods Transparency and Openness. All experimental procedures and materials were reviewed by the National Institute of Pharmacy and Nutrition (OGYÉI; permission number: OGYÉI/22780/2019), and all participants gave informed written consent. All research was conducted in conformity with the Declaration of Helsinki. All materials, all deidentified data on which the study conclusions are based, and all analytic code needed to reproduce analyses are available online and access to the link is provided in Data, Materials, and Software Availability. All fMRI preprocessing and analysis were conducted in the FMRIB Software Library (FSL) (108, 109). Hypothesis testing with metrics derived from the fMRI models and using behavioral data were conducted in R programming language (version: 4.4.3) (110) with the use of the “tidyverse” package (version: 2.0.0) (111) for data organization, the “psycho” package (version: 0.6.1) (112) for calculating d′ values, the “rstatix” package for ANOVA (version: 0.7.2) (113), and “ggplot2” for data visualization (version: 3.5.2) (114). We used GitHub Copilot (115) for code autocompletion, commenting and troubleshooting, and ChatGPT (116) for language editing. The data presented were analyzed via services provided by the HUN-REN Cloud (https://science-cloud.hu/) (117). Participants. Sample size was determined via a power analysis conducted in G-Power 3 (118), based on Nash et al. (35) who detected an interaction of hippocampal region and MST conditions on participant-specific beta estimates (F(51,2,397) = 6.57, P < 0.001, η2 = 0.06) for a power of 0.8 at P < 0.05, yielding a sample size of 30. Between April 2023 and March 2024, 39 young adults (Mage = 21.26, SDage = 1.85, 18 females, all Caucasian) took part in the experiment in exchange for course credits or vouchers at our lab (Brain Imaging Centre, HUN-REN Research Centre for Natural Sciences, Budapest, Hungary) in an urban area. We excluded nine participants during data collection for the following reasons: two for excessive motion during the experiment, one for sudden onset of panic in the scanner, one for left-handedness (due to an error during recruitment), one for incidental finding of brain abnormality, three for missing MRI onsets due to experimenter error, and one with a 56% false alarm rate when recognizing novel items signifying a possible misunderstanding of the task. Thus, the final sample consisted of 30 participants (Mage = 21, SDage = 1.58, 15 females). Participants were recruited through a course at Eötvös Loránd University (ELTE), Budapest, Hungary, and through social media with the help of student research assistants. We selected participants who were right-handed, had no history of neurological, psychological, or cardiovascular diagnosis, had unimpaired or corrected vision under ±6 diopters, as well as participants who were MRI compatible (e.g., had no metal implants, were not claustrophobic, etc.). Behavioral Methods Materials. Main task (sMST). We used the sMST (54) developed based on the incidental two-phase version of the MST (19, 119). In the latter, participants first encode images of everyday objects and later, in a surprise recognition test, make mnemonic decisions about repeated, novel and similar lure images. The sMST uses a set of adjective-noun phrases (e.g. exotic zoo) as stimuli instead of images and similar lures (e.g., strange zoo) consist of a shared concept noun preceded by a semantically similar feature emphasized to the encoded trace. This is analogous to the MST, where similar lures show the same object with similar perceptual features (e.g., a red apple, and a reddish green apple). Importantly, relative to Ilyés et al. (54), here we presented exact repeats and modified repeats during the encoding phase and also presented similar lures during the recognition phase, to be able to measure pattern separation signals during both encoding and recognition (55). Stimuli. We created 225 adjective-noun phrases selected from the “press” subcorpus of the Hungarian National Corpus (HNC) (120), normed and assessed for frequency, character and syllable count, and psycholinguistic measures known to affect recognition memory [concreteness (121), imageability (122), arousal (123), and meaningfulness (124)]. Of these, 100 phrases were used to create 100 triplets comprising the original phrase (e.g. exotic zoo), a semantically close modified (e.g., strange zoo), and a semantically distant modified version (e.g., “alternative zoo”), whereas 125 further phrases were either fillers at encoding or foils during recognition. For parametric manipulation of similarity, we used cosine similarity (A·B∥A∥∥B∥, where the dot product of A and B word vectors is divided by the product of their L2-norms [i.e., the square root of the sum of the squares of a vector’s components]) as a similarity metric, comparing the angle of 200-dimensional word vectors in the Hungarian word2vec (58) word-embedding (57) based on the press subcorpus of the HNC (120)] (for more details, see SI Appendix, Supplementary Methods). Crucially, we have additionally refined the sMST (54) to ensure that both close and distant versions are meaningfully related to baseline phrases. For example, in the study by Ilyés et al. (54), the distant version “public zoo” did not have a meaningful connection to its baseline version exotic zoo (because “public” is not a quality of “being different” like “exotic” is), thus we replaced it with alternative zoo (better capturing the shared meaning of being different). In the implementation used here, we have made replacements to ensure that the three phrases of each 100 baseline—close—distant triplets are meaningfully similar to each other. This assessment was based on independent agreement of three experimenters (A.I., B.M.B., and a third experimenter), because cosine similarity by itself quantifies semantic similarity even between agrammatical or unnatural phrases—and the original list generated using automated procedures included such phrases. In the current version of the design, the difference between close and distant variations is not confounded with the meaningfulness of the similarity itself. Design and procedure. We used a within-subject design for the main task with stimulus condition as an independent variable. The task had two phases, each with separate stimulus conditions: an incidental encoding phase and a two-alternative forced choice recognition phase. Because we presented semantically modified versions of the same phrases both at encoding and recognition, we use different terminology to describe the conditions 1) at encoding: first presentations (of the baseline), exact repeats (of the baseline), modified (close or distant versions) repeats, fillers, and 2) at recognition: targets (repeated baseline phrases), lures (close or distant versions), foils (entirely novel phrases). Both phases were administered inside the scanner, with a 15-min break between them. The encoding phase consisted of three runs. Participants viewed a total of 275 adjective-noun phrases (n = 93 in the first run, otherwise n = 91). There were a total of 175 first presentations (n = 59 in the first run, otherwise n = 58). Of these, 50 were exactly repeated (n = 16 in the second run, otherwise n = 17), 25 had its close modified repeat shown (n = 9 in the second run, otherwise n = 8), 25 had its distant modified repeat shown (n = 9 in the first run, otherwise n = 8), and 75 were not reintroduced in any way (n = 25 in each run). We made sure that both exact and modified repeats of phrases only occurred after the first presentation of the baseline version within a range of 19 to 24 trials (uniformly distributed among conditions within the same runs). To accommodate task instructions to the vocabulary derived from a press corpus, participants were asked to act as editors of a magazine and decide for each phrase whether they would accept or reject it before going in print. This cover story also ensured that participants focused on reading and processing the semantic content of the phrases. In each trial, a fixation cross appeared in the middle of the screen for an interval of either 2, 4, 6, 8, or 10 s (pseudorandomized in each condition to follow a skewed Gaussian distribution with a peak at 4 s). Then, an adjective-noun phrase was presented in the middle of the screen (Open Sans, 81pt) for two seconds with available response options (via two handheld controllers) of “left index finger—would not accept the phrase” written in the bottom left and “right index finger—would accept this phrase” in the bottom right corner of the screen. The recognition phase consisted of two runs. Participants viewed a total of 150 adjective-noun phrases (n = 75 per run). These were either targets (i.e., exact repeats of first presentations at encoding; n = 25 per run), lures (close or distant modified repeats of first presentations at encoding; n = 25), or novel foils (neither the adjective nor the noun seen during encoding; n = 25). Lures were either “close lures” (n = 12 or n = 13, counterbalanced) or “distant lures” (n = 12 or n = 13, counterbalanced) as a function of their semantic distance from their target (Stimuli). Participants were instructed to decide whether they had seen the exact same phrase during the encoding phase. In each trial, a fixation cross appeared in the middle of the screen for a pseudorandomly jittered interval (same as in the encoding phase), followed by a phrase presented on the middle of the screen (Open Sans, 81pt) for two seconds with available response options of “left index finger—not seen” written in the bottom left and “right index finger—seen” in the bottom right corner of the screen. Crucially, to ensure that the correct response for lures was “new,” we made sure that the lures presented during recognition were not shown during the encoding phase, thus half of the recognition lures were presented as exact repeats during encoding, while the other half were presented as a modified repeat during encoding, such that phrases from the previously unseen similarity bin were presented (a “distant lure” was presented at recognition if a “close modified repeat” was presented at encoding, and vice versa). Participants practiced six trials before both phases both inside and outside the scanner. We specifically made sure that they understood that the old response in recognition is only correct if they saw the entire phrase during encoding. Randomization. We generated a single pseudorandomized structure of conditions ensuring that a maximum of three consecutive trials could be from identical stimulus conditions for both the encoding and the recognition phase. This structure was held constant for all participants to allow for comparing participant-wise fMRI data. To avoid confounds arising from the presentation order of the phrases during encoding, for each participant, we assigned one of eight unique lists, in which different phrases were assigned to the conditions counterbalanced such that all phrases appeared equally in both semantically similar versions. For the recognition phase, each participant received unique lists, circumventing list order effects on recognition. Covariate assessments. One month after completing the sMST, participants came back to the lab for a second session (data not analyzed in this article) and completed covariate tasks and questionnaires controlling for confound variables known to affect recognition memory of words. Specifically, we used two subtests from the Hungarian Wechsler Adult Intelligence Scale (WAIS-IV) (125) to assess processing speed (Digit-Symbol Substitution Test) and crystallized knowledge (Vocabulary Test), and administered a Hungarian nonword repetition task (126) to assess phonological awareness and verbal working memory (more details in SI Appendix, Supplementary Methods 1). Analysis. As a proxy for behavioral mnemonic discrimination performance, we calculated the two-choice version of the “lure discrimination index” (LDI) (119, 127): a d-prime (d′) score [calculated as P(old|target) – P(old|lure)] that described the discriminability of responding to lures from responding to repeated targets. Lower d′ means a lower discriminability due to excessive false alarms for specific conditions. Since, the d′ score is an aggregated index across participants, it is not sensitive to fine-grained item-level similarity between recognition lures and encoding first presentations, thus, to assess fine-grained differences of mnemonic discrimination performance as a function of the continuous semantic similarity metric (cosine similarity between phrase vectors, detailed in Stimuli), we also analyzed the probability of responding with old in a GLMEM. We modeled LDI performance via an ANCOVA and we removed one participant from the sample pool for not having provided the participant-level covariates. For transparency, we provide the analyses without the inclusion of covariates (SI Appendix, Supplementary Results 5). For the fine-grained analysis, we ran two GLMEMs predicting response (coded as 0 for new and 1 for old) as a function of cosine similarity in a three-way interaction with lure type (close or distant) and repetition type at encoding (exact repeat or modified repeat), and covariates of vocabulary, processing speed, and phonological awareness (participant-level covariates), as well as meaningfulness, concreteness, and arousal (item-level covariates) in interaction with semantic similarity. In a “baseline” model we only modeled the experimental effects, and in a “covariate” model, we also included fixed effects of vocabulary, processing speed, and phonological awareness as participant-level covariates as well as arousal, meaningfulness and concreteness as item-level covariates. We selected the covariate model with the lower Akaike Information Criterion (AIC) based on a significant difference between model fits (χ2(9) = 18.84, P = 0.03). In this model one participant was excluded due to missing covariates. We reported estimates of the GLMEM that index the effects of contrasts on the likelihood of responding old with an intercept of close lures and being a modified repeat during encoding, such that an estimate >1 signifies a positive change in likelihood, while an estimate <1 signifies a negative change. We also reported CI for these estimates, using the “confint” command of the stats package in R with the Wald method, calculating a 95% CI by z-scoring the 97.5th percentile in both directions and then adding it to the estimate multiplied by the SE reported. MRI Methods. MRI scans were run on a Siemens Prisma 3T scanner with a 32-channel head coil. Functional data were collected with a partial field-of-view multiband EPI sequence [1.5 mm isotropic, TR = 1,000 ms, TE = 32 ms, multiband factor = 6, 60 slices, following Kok et al. (128)] tilted parallel to the long axis of the right HC, covering the most dorsal point of the occipital lobe. High-resolution T1-weighted structural images of the whole-brain were acquired at the beginning of the experiment using an MPRAGE sequence (1 mm isotropic, TR = 2,300 ms, TE = 3.03 ms, 192 slices). High-resolution proton-density (PD) images covering the medial temporal lobe were acquired after the short interim break in the task (0.4 × 0.4 × 2 mm, TR = 649 ms, TE = 16 ms, 44 slices), following Keresztes et al. (129) and Bender et al. (130) tilted perpendicular to the right hippocampus. Preprocessing. First, we discarded the initial six volumes of each functional scan to ensure a steady state of acquisition. We prefiltered each scan using a 100 Hz temporal filter (calculated by FSL on the custom design matrices) to remove low frequency drifts during scanning, then we used the MCFLIRT tool (131) for motion correction and to derive six motion regressors that were introduced in later modeling. Then, we removed nonbrain tissues from T1 images using FSL’s Brain Extraction Tool (BET) (132). Finally, using FMRIB’s Linear Image Registration Tool (FLIRT) tool (133), we registered each functional scan to the participant’s T1 image at 6 degrees of freedom (DoF). During this step, we also used phase-magnitude fieldmaps for the encoding runs and AP/PA fieldmaps for the recognition runs acquired before each task phase to correct for inhomogeneities of the magnetic field. Before first-level analysis, to enhance signal-to-noise in the standard space, we smoothed the functional images using a 2-mm Gaussian kernel. For group-level analysis, we additionally registered all data to a sample-specific template brain in MNI space (T1 template) that we created using the T1 images via “antsMultivariateTemplateConstruction2” from the Advanced Normalized Tools (ANTS) (134). Hippocampal subfield segmentation. We used two separate protocols to segment the group level hippocampal body and the hippocampal head/tail. For the body, we used a state-of-the-art harmonized protocol (103). For the head, since no harmonized protocol was available, we reverted to using an openly available atlas (135) for a similar acquisition as ours based on a protocol (136) incorporating anatomical discrepancies for delineating the anterior HC. We first created a sample-specific template image from high-resolution PD images as well (PD template) using “antsMultivariateTemplateConstruction2.” Two trained researchers (A.I. and A.K.) agreed on the anterior and posterior boundaries (i.e., ranging) of the hippocampal body on the template image following Bender et al. (130), having reached an interrater reliability of Cohen’s κ > 0.62 on each boundary in the single PDs of the total sample via the same guideline. First, subfields in the HC body were segmented manually by A.K. on the PD template—having extensive (>10 y) experience with manual tracing—based on a harmonized protocol (103). We then registered these masks to the T1 template using ANTS (SI Appendix, Fig. S10). We specifically used a two-step rigid (3 DoF) and affine (6 DoF) registration pipeline, whereby each step was run in two stages: First, we administered a whole-brain registration followed by a refined, more conservative registration using masks of the medial temporal lobe. Lacking a harmonized protocol for the hippocampal head, we first automatically segmented the HC of each participant with ASHS (137) using an atlas with subfields defined for the hippocampal head in a similar scanning sequence to ours (135, 136). We then registered individual masks to the individual T1 acquisitions. To enhance group-level alignment, we then used the registration transformation matrices from the analysis pipeline to register each hippocampal head mask in T1 space to the T1 template, then we selected a subfield label for each individual voxel only if 75% of participants had an overlapping subfield label there (with the majority deciding on the subfield identity). Using these two segmentation methods, we obtained masks of the following subfields in both the hippocampal head and body: cornu ammonis 1-2 (CA1-2), dentate gyrus-CA3 (DG-CA3), and subiculum (all coronal slices of the T1 template with HC masks included are available in SI Appendix, Supplementary Methods 2). Note that both protocols separately segmented the DG, CA1, CA2, and CA3, but we decided to merge the DG and CA3, and the CA1 and CA2 subfields according to conventions into DG-CA3 and CA1-2 region of interests, because in the template T1 space the CA2 mask spanned 8 voxels and the CA3 mask spanned 44 voxels, which would have caused the anatomical control analyses to have low signal-to-noise ratio and reliability. Analysis. We used generalized linear modeling (GLM) with FSL’s FEAT tool (138) to estimate regressors for each stimulus condition in both phases of the main task. First, we defined and ran models within each functional run (“first level”) of all participants using FMRIB’s local analysis of mixed effects (FLAME). Second, we ran fixed effect models averaging across runs in both task phases for each participant (“second level”). Finally, we ran mixed effect models across participants (“group level”). In the encoding phase, to detect repetition sensitive clusters in the HC at the group level, we created an “aggregated model” with regressors for exact repeats, modified repeats, and first presentations binned as one to mitigate statistical circularity as the second level “separated model,” used for hypothesis testing, contained regressors each first presentation (exact, close, distant) and modified repeat (close, distant) separately. We then thresholded the parameter estimate of the critical first presentation—exact repeat contrast (signifying RS) at the group level with an uncorrected P < 0.05 and a spatial extent of at least 50 contiguous voxels (>75 mm3). This liberal threshold was selected in line with previous research to uncover functionally relevant voxels while reducing voxel-selection bias (139). In another separated model, we specified regressors for first presentations separately for each repeat type, exact repeats, close modified repeats, distant modified repeats, and fillers to use for extracting participant-specific percent signal changes for data analysis. In the recognition phase, we defined a “separated lures model” with regressors of target, close lure, distant lure, and foil split based on memory correctness (correct and incorrect modeled separately). In the recognition model, we binned trials together based on the type of repeat at encoding (exact repeat or modified repeat) to maximize the number of trials in all correct and incorrect bins in each run. For a follow-up exploratory analysis, we also made an “encoding-based recognition model” such that we defined regressors separately for lure correct rejections and target hits based on repeat type at encoding (SI Appendix, Supplementary Results 7). We also binned trials that received no response in the given timeframe with the incorrect trials, as we observed no successful mnemonic discrimination. In all models, interstimulus-intervals functioned as an implicit baseline during modeling. We used a double-gamma hemodynamic response function (HRF) to convolve BOLD responses from voxel time series. We extracted parameter estimates converted to percent signal changes using “Featquery” (140) from second level GLMs using the clusters detected at the group level to model via hypothesis-specific ANCOVAs. To assess the specificity of RS during encoding, we modeled second-level percent signal changes with fixed effects of condition (exact repeat or modified repeat), order (first or second presentations), and cluster (left or right hippocampal head) in interaction. To assess differences in signal during recognition, we modeled second-level percent signal changes with fixed effects of condition (target, lure, foil) and cluster (left or right hippocampal head). In both models we included the participant covariates that showed a significant effect during recognition performance (processing speed and phonological awareness). Due to missing covariate assessments, we excluded one participant from these analyses. To ensure normality of distribution in each fixed effect-based bin during the analysis of the encoding data, we also excluded one outlier using a three interquartile range above and below the third and first quartiles of each bin (using the “identify_outliers” function of the rstatix). Furthermore, for the correlation analysis we also excluded the same outlier (detected by two separate measures) based on a multivariate normal distribution via Mahalanobis distance (MD) with a cut-off at χ2(2) = 0.99. For transparency, we provide the analyses without the inclusion of covariates and the influential outlier filtering (SI Appendix, Supplementary Results 6). Overall Procedure. Participants arrived at the lab and signed an informed written consent. We made sure that participants were comfortable with going inside the MR, and we also presented the first task phase and administered a short practice before going inside. They then completed an MR screening to make sure that no metal compromising safety or data quality entered the scanner room. We performed MRI acquisitions in the following order: T1 anatomical scan (~5 min) before the encoding phase, which was repeated upon excessive movements, fieldmap and the three encoding functional runs (~12 min each), PD anatomical scan (~4 min), which was repeated upon excessive movements and finally fieldmap and the two recognition runs (~10 min each). For a 15-min self-paced break in-between the two task phases, participants left the scanner room and completed the practice for the recognition phase. Participants received course credits or remuneration (6,000 HUF ~ 15.5 EUR) after completing the session. Supplementary Material Appendix 01 (PDF) Dataset S01 (XLSX)