Eight patients with drug-resistant epilepsy implanted with depth electrodes in the hippocampus for presurgical evaluation of stereo-EEG participated in the study. Two patients were excluded due to the presence of persistent interictal spikes in the recordings. From the remaining patients, two retrieval sessions were excluded from the analysis due to the presence of interictal epileptic discharges and artifacts in the 100% of each recorded session. In total, the analysis included 6 patients for the encoding session (mean age = 30.5, S.E. = 5.6, 5 females) and four patients (mean age = 32, S.E. = 8.6, 3 females) for the retrieval session. All patients were right-handed, had normal vision and provided written informed consent as approved by the institutional ethics review board (PB 2016–02055). The sample size is in line with the sample sizes of other stereo-EEG studies with similar long-term memory paradigms (see Axmacher et al., 2010; Castelhano et al., 2022; Mormann et al., 2007; Staresina et al., 2012; Sweeney-Reed et al., 2016; Vila-Vidal et al., 2023).

Each participant attended two sessions: encoding and retrieval over two consecutive days. During the encoding session, participants performed an associative task which included an estimation of congruence of encoded material (object-scene pairs). On the next day, during the retrieval session, participants performed recognition and associative memory tests. The mean intersession gap was 23 ± 1 h.

The associative memory task was adapted from van Kesteren et al. (2013). The set consisted of 185 pairs of pictures (one representing an object and one representing a scene). Pictures were recognizable colored photographs representing indoor and outdoor scenes and objects of different sizes. Each object-scene associative pair was unique (i.e., every picture was presented only once to each participant). The pairs were initially constructed to be very congruent (e.g., “book” and “library,” 10% of pairs), of medium congruence (e.g., “earplugs” and “living room,” 80% of pairs) or very incongruent (e.g., “lab” and “beachball,” 10% of pairs). On the first day, participants were presented with a set of 185 object-scene pairs. They received instructions to remember “everything they see on the screen.” Following the presentation of an object-scene pair, the participants estimated the semantic congruence of the object-scene pair using a visual analogue scale ranging from “does not fit” to “fits very well.” Their answers were referred to the congruence of the stimuli.

The trial structure is shown in Figure 1A. Before stimulus onset a black fixation cross on a gray background was presented for 350 ms, followed by a gray blank screen for 100 ms. The visual stimulus, composed by a pair of pictures representing an object and a scene, was presented for 2.5 s. The object was presented on the right and the scene was presented on the left of the screen. After 100 ms of blank screen, a visual analogue horizontal scale appeared. Participants had unlimited time to answer the question “how well does this object fit this scene?” moving a pointer along the visual analogue scale with a computer mouse. After the congruence rating, the screen remained blank for 1.5 s before the start of the next trial.

On the second day, participants completed the retrieval session, where they were presented with 300 objects (185 old +115 new) and asked to distinguish between familiar items learned in the encoding part and new items (item recognition test). The old and new items were presented in random order. Following the answer “new,” the next item was presented. Following the answer “old,” the participant was asked to remember in which context this item appeared previously and to choose from three options (associative memory test). Options were presented as one or two-word verbal descriptions to achieve more categorical-type recognition than perceptual features recognition (van Kesteren et al., 2013). Finally, the participants were asked how sure they were about their response (“guess,” “not sure,” “very sure”). Participants had 6 s to answer each question. The experiment timeline is shown in Figure 1A.

Multiple-choice options for the associative memory test were manually predefined so that, together with the correct answer, they included two congruent options and one incongruent option (van Kesteren et al., 2013).

The paradigm was implemented in E-Prime (version 2.0.10.147, Psychology Software Tools, Pittsburgh, PA).

The recording was carried out in the intensive monitoring unit of the Swiss Epilepsy Center in Zürich using electrodes with 8 recording contacts (diameter 1.3 mm, contact length 1.6 mm AD-Tech, www.adtechmedical.com). Stereo-EEG was recorded with an ATLAS recording system (sampling rate 4 kHz, 0.5–1,000 Hz bandpass, Neuralynx, www.neuralynx.com).

Localization and anatomical labeling of contacts was performed using the protocol described in Stolk et al. (2018) based on the Brainnetome Atlas (Fan et al., 2016) after merging the post-operative MR with post-operative CT images in MNI space. The neurosurgeon (LS) confirmed the anatomical labels of the contacts in the participants’ native space (iPlan Stereotaxy 3.0, Brainlab, Germany). Contacts located in the HC and their distribution across patients are shown in Figure 1B.

The stereo-EEG signal was downsampled to 200 Hz. We adopted Common Average Reference in agreement with previous studies on stereoEEG data (Arnulfo et al., 2015; Mercier et al., 2022). The data was visually inspected in order to remove episodes that included epileptic spikes and technical artifacts. After preprocessing, trial rejection rate was equal to 26% (SEM = 8%) for the encoding data and 28% (SEM = 9%) for the retrieval data across participants. While the noise level may vary across two days of recordings, we compared the percentage of rejected trials across sessions and found no significant difference [t(7.04) = 0.12, p = 0.91]. Data was then epoched from −2 to 3 s around the stimulus onset. Pre-processing of the stereo-EEG signal was performed in Brainstorm (Tadel et al., 2011).

Firstly, we estimated the distribution of items classified by the subjects as congruent, intermediate, and incongruent (Figure 2B). A Friedman rank sum test with congruence as a factor did not reveal significant differences in their amount: χ²(2) = 2.33, p = 0.311. The Spearman correlations between individual subjective and objective congruence rates are presented in Supplementary Table 2 (mean R² = 0.27, all p-values <0.001, showing a significant correlation between subjective and objective rates).

Secondly, based on the memory performance of the participants, the encoding and retrieval trials were classified into four categories (see Materials and Methods and Figure 2A): hits, misses, correctly associated items (CA) and incorrectly associated items (IA). Behavioral performance in terms of recognition memory and associative memory in the recorded patients, though lower than in van Kesteren et al. (2013) study, was in line with the expected performance in healthy individuals (Snodgrass and Corwin, 1988). Subjects’ individual performance metrics are presented in Table 1. Recognition memory performance, provided by a discrimination index (DI), was significantly above chance level (0) [t(5) = 10.91, p < 0.001, d = 4.46]. Associative memory performance, measured as a proportion of correctly associated items, was above chance level (0.33) [t(5) = 4.63, p = 0.006, d = 1.89]. Mean (SEM) values of memory performance are shown in Table 1.

Finally, we assessed the distribution of correct answers (Hits and CA) in every congruence bin. The percentage of hits at each congruence level is represented in Figure 2C, while the percentage of CA is shown in Figure 2D. A Friedman rank sum test with congruence as a factor did not reveal significant differences between the amount of correctly recognized items of each congruence level [χ²(2) = 2.33, p = 0.311]. However, a Friedman rank sum test with congruence as a factor showed significant differences in the amount of correctly associated items at each congruence level [χ²(2) = 7.00, p = 0.030]. Pairwise Wilcoxon rank-sum tests were conducted to compare the levels of CA across the congruence conditions. The results, adjusted using the Bonferroni correction, revealed a significant difference between the congruent and incongruent conditions (p = 0.026). No significant differences were found between the congruent and intermediate conditions (p = 0.195) or between the incongruent and intermediate conditions (p = 0.195). Therefore, associative memory performance was significantly higher for congruent than incongruent items, indicating that congruent object-scene pairs were more likely to be correctly recalled.

We quantified the modulation in the time-frequency domain in relation to recognition memory and associative memory performance and subjective congruence estimation. We tested the recognition memory effect through a mixed-effects GLM with oscillatory power as a dependent variable and “Recognition memory” (“Hits” and “Misses”) and “Congruence” (“Congruent,” “Intermediate,” and “Incongruent”) as explanatory categorical variables. We tested the associative memory effect through a mixed-effects GLM with oscillatory power as a dependent variable and “Associative memory” (“CA” and “IA”) and “Congruence” (“Congruent,” “Intermediate,” and “Incongruent”) as explanatory categorical variables. This framework was applied to rostral and caudal HC during the encoding (day 1) and retrieval (day 2) sessions separately.

For a complete overview of the GLM models, direct comparison of trials sorted by memory performance and direct comparison of trials sorted by congruence rating (see Supplementary Figures 1, 2).

At encoding, our generalized linear model (GLM) analysis did not reveal any significant Subsequent Memory Effect (SME) for either the rostral or caudal portions of the hippocampus (Supplementary Figures 1A,B, top left panels). This finding aligns with established evidence suggesting that hippocampus-independent processes support the encoding of items that are later correctly recognized (Ranganath et al., 2004; Staresina and Davachi, 2006).

During retrieval, however, correct item recognition was specifically associated with an early alpha band power decrease within the rostral HC (Figures 3A,B, top right panel). Recognition involves both perceptual processing of the stimulus and a familiarity-based matching-to-sample search. The observed alpha suppression likely reflects the inhibition of low-frequency idling rhythms, facilitating information exchange between the HC and cortical regions. This is consistent with previous findings that link alpha activity suppression to resource allocation during memory and attention-related tasks (Klimesch, 2012; Klimesch et al., 2007). In this context, the alpha power decrease could represent an adaptive mechanism for engaging strategic memory search (Griffiths et al., 2019; Iemi et al., 2022).

Following the alpha suppression, we observed a subsequent gamma power increase in the rostral HC (Figures 3A,B). Gamma oscillations have been associated with the formation of large-scale cortical memory networks, which temporarily integrate distributed neural representations (Busch et al., 2006; Engel et al., 1990; Jacobs and Kahana, 2009; Martinovic et al., 2007). Interestingly, our findings reveal that this gamma activation occurs approximately 0.5 s after stimulus presentation and follows the alpha suppression phase. Previous research has shown that familiarity-driven cortical gamma activity often precedes hippocampal gamma responses (Supp et al., 2007). In this respect, the gamma power increase observed during retrieval likely reflects not only re-encoding of the stimulus or reactivation of memory traces but may also contribute actively to the computational processes underlying recognition.

During encoding, the rostral HC showed a selective Subsequent Memory Effect (SME), characterized by an increase in gamma power for correctly associated items compared to incorrectly associated items and misses (Figure 3C,D). Gamma power increases during encoding are widely recognized as key mechanisms for item memory formation and episodic binding, facilitating the integration of individual elements into cohesive memory traces (Hanslmayr et al., 2016; Headley and Weinberger, 2011; Henin et al., 2019; Sederberg et al., 2003; Staresina et al., 2016). This process appears to be particularly associated with high-frequency gamma activity, which has been linked to local processing and network coordination during memory encoding (Burke et al., 2014; Griffiths et al., 2019; Jobst and Cascino, 2015; Tort et al., 2009).

At retrieval, successful associative memory recall was accompanied by a distinct pattern of neural activity in the caudal HC, with concurrent decreases in low-frequency power and increases in high gamma power (Figures 4C,D). These findings build upon previous evidence linking hippocampal activity to associative memory retrieval (Griffiths et al., 2019; Staresina et al., 2016), but our results uniquely highlight the specific engagement of the caudal HC during this process. The combined modulation of low- and high-frequency power may reflect a dual mechanism: low-frequency power decreases likely facilitate inter-regional communication by reducing inhibitory rhythms, while gamma power increases are associated with local computations required for memory retrieval.

During encoding, caudal hippocampus (HC) activity exhibited a greater decrease in low-frequency power for congruent and incongruent items compared to intermediate ones (Figures 4A,B). Concurrently, a similar pattern was observed during the retrieval of associative memory (Figures 4E,F). This finding highlights the role of the caudal HC in associative encoding and provides new evidence on how congruence modulates hippocampal dynamics. While ripple activity is strongly associated with memory encoding and is thought to be predominantly driven by the caudal HC (Jiang et al., 2019), our results complement this framework by demonstrating that low-frequency power suppression in the caudal HC is also sensitive to the congruence of associative information.

Congruence is a critical factor in memory processing, as it determines the compatibility of new information with pre-existing knowledge structures, such as schemas. According to the theories proposed to explain schema-based processing, congruent information is processed more efficiently because it integrates seamlessly into existing frameworks, while incongruent information requires additional resources for encoding and integration (Ghosh and Gilboa, 2014; van Kesteren et al., 2012). The observed low-frequency power decreases for both congruent and incongruent items suggest that the caudal HC plays a dual role in encoding: facilitating rapid integration of congruent information while also adapting to the demands of encoding incongruent inputs. This dual role underscores the flexibility of the caudal HC in mediating familiarity-driven and novelty-driven processes. A similar pattern is observed during retrieval, reflecting the continuity of underlying neural mechanisms across encoding and retrieval. Strongly congruent and strongly incongruent items are retrieved more easily due to their clear fit or distinct separation from existing schemas. In contrast, items with moderate congruence demand greater cognitive effort as new associations must be established.

The encoding of associative material is accompanied by an increase in hippocampal (HC) high gamma power (Henin et al., 2019), which may reflect the integration of items into cohesive memory representations. Successful retrieval (recognition and recall), on the other hand, has been associated with the modulation of theta (Anderson et al., 2010; Foster et al., 2012; Sederberg et al., 2006) and gamma rhythms (Sederberg et al., 2007; Steinvorth et al., 2010), and their interaction (Foster et al., 2013; Mormann et al., 2005). Theta rhythms are thought to coordinate the activity of distant brain structures during memory retrieval, while gamma oscillations provide a foundation for localized computations and integration (Knight and Eichenbaum, 2013; Von Stein and Sarnthein, 2000; Watrous et al., 2013).

Our findings contribute to the understanding of the HC’s role in associative memory by providing evidence for distinct spatial and functional activation patterns along the rostrocaudal axis. Initially, the HC was hypothesized to function primarily as a novelty detector (Brodt et al., 2016; Rutishauser et al., 2008; Tulving et al., 1996). More recent evidence has demonstrated its involvement in context recall and feature binding, extending its role beyond novelty detection to include object-context integration (Staresina et al., 2013; Staresina and Davachi, 2006).

In our study, we examined HC activity in object-context associations, which inherently involve episodic experiences and semantic knowledge. These processes align with the concept of “predictable ambiguity,” where stimuli can have multiple meanings depending on contextual variations (Morris, 2006). The HC appears to resolve such ambiguity through a dynamic interaction of gamma rhythms that segregate parallel computational processes. Specifically, at retrieval, we observed distinct patterns of HC gamma activity: higher gamma activity in the caudal HC during associative memory retrieval and lower gamma activity in the rostral HC during item recognition. This functional dissociation supports the hypothesis that gamma oscillations can disentangle interfering computational processes (Colgin et al., 2009) coding separately for novel and familiar information. Presumably, variability in the gamma spectral profile could code specifically for novel and well-known information. Slow gamma oscillations are supposed to be driven by intrinsic HC pacemakers, while fast gamma oscillations are thought to originate from the medial entorhinal cortex (Bragin et al., 1995; Colgin et al., 2009). Interneuronal activity within the HC may mediate the competitive relationship between these states, enabling the flexible encoding and retrieval of associative information (Colgin, 2015; Leão et al., 2012).

To further clarify the spatial and temporal dynamics of the information flow from and toward the HC, future research should incorporate concomitant extrahippocampal sampling. Evidence from fMRI studies has identified two functional subnets within the memory network: an early subnet associated with high-level visual perception and a later one linked to top-down control mechanisms (Buckner and Koutstaal, 1998). However, the detailed time-frequency profile of these subnets remains poorly understood. Current evidence suggests that memory processes involve gamma activity increases in brain regions adjacent to the HC, including the temporal cortices (Fell et al., 2001, 2003; Matsumoto et al., 2013), medial prefrontal cortex (Milivojevic et al., 2015; van Kesteren et al., 2010, 2013), inferior frontal gyrus (Burke et al., 2014; Hoffman et al., 2015; Long et al., 2014), angular gyrus (Davis et al., 2020; Davis and Yee, 2019), precuneus (Brodt et al., 2016) and posterior parietal cortex (Long et al., 2014). Our findings suggest that rostrocaudal HC activity patterns contribute to understanding the mechanisms underlying the associative memory network. By characterizing the differential roles of the rostral and caudal HC, we offer insights into how distinct hippocampal regions coordinate with broader cortical networks to support associative memory retrieval.

The anterior (rostral) and posterior (caudal) partitions of the human hippocampus (HC) play distinct roles in memory function (Strange et al., 2014). The HERNET model, based on a meta-analysis of fMRI memory studies, suggests that the rostral HC is predominantly associated with the dorsal attentional network, supporting external information processing, while the caudal HC is linked to internal-oriented attention processes mediated by the default mode network (Fritch et al., 2020; Kim, 2015). For example, the two partitions are actively but differentially involved in working memory processing (Grady, 2020), which relies on rostrocaudal information flow (Li et al., 2021). Additionally, the caudal HC has been shown to play a critical role in the recall of verbal information (Lin et al., 2019). Our findings align with the predictions of the HERNET model, demonstrating that successful encoding of episodic information is supported by rostral HC activity, while episodic retrieval relies on caudal HC activity. At retrieval, the caudal HC is specifically involved in recollection rather than familiarity-based item recognition, consistent with the model’s predictions (Kim, 2015). This differentiation suggests that associative recall engages more complex, semantic-based mechanisms that are constructive in nature, whereas visual recognition primarily involves re-encoding and comparative processes.

The functional distinction we observed supports the idea that the rostral HC is optimized for the integration of external sensory inputs during encoding, facilitating episodic memory formation. In contrast, the caudal HC appears to support the retrieval of associative information by reconstructing contextual and semantic details stored across memory networks. This rostrocaudal segregation reflects the broader organizational principles of the HC, where anterior and posterior regions are specialized for distinct attentional and mnemonic functions. Our findings further emphasize the importance of understanding these differential roles to fully characterize how the HC contributes to complex memory processes.

The current study provides evidence of the multifaceted involvement of the human HC in memory processing. However, some limitations should be considered. First, the current study has a low sample size (N = 6). However, this limitation is typical for stereo-EEG studies since they are strictly constrained by the capacity of surgical centers and high data dropout due to patient-specific clinical factors (Youngerman et al., 2019). Despite the small cohort, comparable sample sizes have been employed in other stereo-EEG studies that have provided critical insights into memory processing (Axmacher et al., 2009, 2010; Castelhano et al., 2022; Mormann et al., 2007; Staresina et al., 2012; Sweeney-Reed et al., 2016; Vila-Vidal et al., 2023). Importantly, consistent behavioral and neural patterns were observed within our cohort, supporting the robustness of our findings despite the sample size limitation. Second, confidence rates were not analyzed because the distribution of responses across different categories was highly uneven. Specifically, certain combinations of congruence and confidence levels were underrepresented (e.g., trials with High congruence but Low confidence accounted for only about 4% of all trials). This limited data in certain categories made it difficult to conduct a meaningful statistical analysis, as the sample sizes were too small to draw reliable conclusions. Besides, due to the low number of high-confidence responses, we cannot completely rule out the possibility that the observed effect is partially influenced by a guessing strategy. However, if the relationship between congruence and associative memory were merely a by-product of guessing (i.e., participants choosing the congruent option when they do not remember well), we would expect accuracy in such case to be below chance level. Finally, the study focused exclusively on the HC, given the spatial constraints of stereo-EEG, which is guided by presurgical clinical hypotheses (Parvizi and Kastner, 2018; Youngerman et al., 2019). While this limited the ability to capture broader hippocampal-cortical interactions, it allowed for a detailed delineation of the functional and anatomical segregation along the rostrocaudal axis of the HC. These findings provide a foundation for future investigations into hippocampal-cortical dynamics and their role in associative memory processing.

While these limitations impose some constraints on the generalizability of the results, the current study contributes critical evidence to the field and highlights the importance of further research to expand our understanding of hippocampal function in associative memory.

The functional segregation of the hippocampus (HC) along its rostrocaudal axis has important implications for presurgical evaluation and the minimization of cognitive loss. Current evaluations of residual hippocampal functionality primarily rely on tomography studies (Montaz-Rosset et al., 2019; Vinton et al., 2007; Wong et al., 2010) and functional imaging protocols such as fMRI (Schacher et al., 2006), which are applicable to non-implanted patients (Buck and Sidhu, 2020; Duncan et al., 2016). While fMRI and EEG studies have demonstrated correspondence for slower time scales (Fedele et al., 2020; Shamshiri et al., 2019), neural dynamics characterized by fast modulations, such as gamma oscillations, are better captured by neurophysiological recordings like stereo-EEG. In addition, lesions can induce significant functional reorganization within the core memory network (Alessio et al., 2013; Benke et al., 2006; Bernhardt et al., 2019; Golby et al., 2002; Tsukiura et al., 2002). Understanding memory-related patterns of rostrocaudal HC activity can provide critical insights for the interpretation of neuropsychological test outcomes, especially in cases where compensatory mechanisms or reorganization might obscure traditional markers of HC function. By clarifying these memory-related activity patterns, our findings can inform surgical planning and enhance rehabilitation protocols for patients with hippocampal lesions or dysfunction. For example, knowledge of rostrocaudal functional specialization can guide surgeons in preserving critical memory-related regions, reducing the risk of cognitive impairment. Moreover, such insights can help design personalized rehabilitation protocols aimed at leveraging residual HC function or enhancing compensatory mechanisms (de Andrade Morange et al., 2022; Wang et al., 2023).

In summary, the study’s contributions to understanding HC functional segregation have the potential to improve both the diagnostic and therapeutic strategies employed in clinical settings, ensuring better patient outcomes and more targeted interventions.