Reassessment of the Hippocampal Index Theory: A Synopsis of the Present Evidence and Computational Models

The connection between how memory is represented in the brain and how it can be computationally modeled lies at the heart of contemporary neuroscience. This review revisits the Hippocampal Index Theory in light of recent studies to investigate the structural and functional organization of memory using computationally inspired models derived from theory. First, we summarize how hippocampal subregions, particularly CA3 and the Dentate Gyrus, share functions through complementary mechanisms such as pattern completion and pattern separation. We then examine hippocampal encoding from a physiological perspective, drawing on recent work on engram cells, episode-specific neurons, and abstraction dynamics. We then evaluate various modeling strategies, including deep learning architectures, generative approaches, hippocampus-inspired systems, and complementary learning frameworks, and assess how well they capture the reconstructive and generalizing behavior of memory. These models are not limited to biological mimicry; rather, they generate testable hypotheses and improve mechanistic explanations of hippocampal function. The synthesis of empirical evidence and modeling studies reveals that advanced machine learning algorithms and artificial intelligence systems can serve as explanatory and fundamental frameworks for complex cognitive processes such as memory. Reassessment of the Hippocampal Index Theory: A Synopsis of the Present Evidence and Computational Models Kaan Cengiz 1, Güldal Süyen 2 1 Department of Neuroscience, Graduate School of Health Sciences, Acıbadem Mehmet Ali Aydınlar University, Türkiye 2 Department of Physiology, School of Medicine, Acıbadem Mehmet Ali Aydınlar University, Türkiye

The connection between how memory is represented in the brain and how it can be computationally modeled lies at the heart of contemporary neuroscience. This review revisits the Hippocampal Index Theory in light of recent studies to investigate the structural and functional organization of memory using computationally inspired models derived from theory. First, we summarize how hippocampal subregions, particularly CA3 and the Dentate Gyrus, share functions through complementary mechanisms such as pattern completion and pattern separation. We then examine hippocampal encoding from a physiological perspective, drawing on recent work on engram cells, episode-specific neurons, and abstraction dynamics. We then evaluate various modeling strategies, including deep learning architectures, generative approaches, hippocampus-inspired systems, and complementary learning frameworks, and assess how well they capture the reconstructive and generalizing behavior of memory. These models are not limited to biological mimicry; rather, they generate testable hypotheses and improve mechanistic explanations of hippocampal function. The synthesis of empirical evidence and modeling studies reveals that advanced machine learning algorithms and artificial intelligence systems can serve as explanatory and fundamental frameworks for complex cognitive processes such as memory.

Memory Recall, Memory Retrieval, Hippocampal Index Theory, Episodic Memory, Computational Models

The enigma underlying the phenomenon of memory remains a subject of profound fascination. At this juncture, contemporary neuroscience and memory research are pursuing two fundamental approaches in parallel to ascertain responses to these inquiries. First, there is the effort to understand the complex and often ”disordered” reality of biological systems. Second, there is the approach of creating elegant yet functional computational models that can explain and mimic this reality. The following inquiries are among the most salient that these perspectives engender: The question of how closely computational models should adhere to biological realism, and to what extent this realism should be compromised for the sake of high computational efficiency and detailed simulations, is a complex one. In essence, the question pertains to the extent of its ability to accurately replicate the complex processes occurring within the human brain. It must be acknowledged that a definitive response to the aforementioned inquiry remains elusive at this juncture. However, this compilation seeks to delve into the foundational principles of memory and contemporary machine learning models that draw inspiration from the mechanisms underpinning these principles. This exploration unfolds from the vantage point of the aforementioned inquiries, offering a multifaceted examination of the intricacies and interconnections between these fundamental concepts and the cutting-edge models that emulate their operations. On the one hand, the field of neuroscience provides a comprehensive set of phenomena that demonstrate how the hippocampus ”reconstructs” an entire memory from a partial cue. Conversely, machine learning and artificial intelligence, drawing inspiration from these biological principles, present computational models that seek to emulate the multifaceted functions of memory. However, these models are constrained by their own distinct objectives and parameters. These models and simulations range from direct architectural replication of a biological circuit to the application of a high-level, specific cognitive principle in a completely different domain (abstract principle engineering). In this compilation, we will address the theoretical underpinnings of the hippocampus, with a particular focus on the extant evidence for fundamental neurobiological mechanisms, such as pattern completion and pattern separation. We will also examine the scientific contradictions that this evidence reveals. Subsequently, an examination will be conducted of various computational approaches inspired by these mechanisms, including deep learning, generative models, and hybrid architectures. The subsequent discussion will address the manner in which these computational models provide solutions to fundamental problems in machine learning and enable the generation of new hypotheses concerning the nature of reconstructive memory. The Mystery of Episodic Memory A particularly intriguing facet of the concept of ”remembering” pertains to the faculty of creating a process that enables the experience of an entire memory from a partial cue. This phenomenon can be conceptualized as the ability to assemble an entire puzzle, which is nearly perfect in its completion, from a limited number of pieces with great speed and precision. The olfactory sensation of a flower, the auditory impression of an old friend, or the initial notes of a familiar song can prompt a memory experience that is remarkably realistic and multidimensional within milliseconds. As Schacter and Addis (2007) have noted, this process is referred to as ”constructive episodic simulation.” This suggests that memory recall is, in fact, a form of mental time travel. The ”constructive episodic simulation” model posits that past experiences are flexibly reassembled to both remember past events and imagine future scenarios. This approach underscores the notion that memory recall is not a deterministic process; rather, it is a creative and occasionally erroneous reconstruction process. It is imperative to ascertain the mechanisms employed by the brain to facilitate this intricate reconstruction process. Index Theory The response to this query prompts a conceptual transformation, wherein the function of the hippocampus transitions from a mere content repository to a mechanism for access. The intellectual underpinnings of this revolution were established by David Marr in 1971, when he proposed that the hippocampus functions as an ”addressing” system, emphasizing not the information itself, but its location in the cortex (Marr, 1971). Marr’s seminal concept has led to the prevailing understanding that the hippocampus functions more akin to an extensive catalog, providing a reference point that facilitates the retrieval of information, rather than a physical library. The seeds that had been sown by Marr sprouted fifteen years later when Teyler and DiScenna (1986) formally formulated the ”Hippocampal Index Theory.” According to this theory, the hippocampus functions as a temporary storage system for memories, not as a repository for the memories themselves. Instead, it serves as a guide, storing the ”addresses” of neocortical representations (e.g., visual, auditory, emotional) that constitute those memories. When an experience occurs, groups of neurons are activated simultaneously in different regions of the cortex. These neurons are linked together by an extremely sparse ”index” created by the hippocampus (Teyler and DiScenna, 1986). Subsequent years have witnessed updates to this theory, thereby deepening its contributions to the understanding of episodic memory (Teyler & Rudy, 2007). The ”sparsity” of the model is of critical importance, as utilizing a limited number of neurons to represent a memory prevents the conflation of similar memories. The process of recalling a memory is initiated by a partial cue, such as a scent or melody. The activated index resynchronizes the dispersed cortical modules that were previously active during the original experience, thereby reviving the entire memory experience. This process occurs through two mechanisms: pattern completion and pattern separation (Hunsaker, 2013; Rolls, 2013; Tanaka and McHugh, 2018; Teyler and DiScenna, 1986; Yassa and Stark, 2011). Pattern Completion and Pattern Separation Pattern completion is defined as the cognitive process of reconstructing a previously stored memory from incomplete or distorted cues. The hippocampus, and particularly the CA3 subregion, are the primary components of this process. CA3 performs this function as an auto-associative network (Yassa and Stark, 2011). Contrary to popular belief, the strength of this network does not stem from dense connections. Guzmán et al. (2016) demonstrated that, despite the sparsity of these connections, they are abundant in disynaptic motifs that facilitate pattern completion. Gershman et al. (2025) proposed that the hippocampus manages access to information through a ”key–value” architecture. The hippocampus’s ”key” representations are defined as addresses that regulate access to ”value” contents stored in the neocortex. Incorrectly activated key representations are corrected by the attractive network dynamics of CA3, resulting in the establishment of the correct memory representation (Gershman et al., 2025). This approach elucidates the classical pattern completion process in terms of a modern addressing system. Conversely, pattern separation is a process that modulates recall effects by orthogonalizing the neural representations of similar experiences. The Dentate Gyrus (DG), the primary mechanism responsible for this phenomenon, plays a critical role in preserving the uniqueness of hippocampal encoding (Bakker et al., 2008; Leutgeb et al., 2007; Yassa et al., 2011). Refer to Figure [1]. The following figure explore the phenomenon of pattern separation and pattern completion in the context of memory processing within the hippocampus. (Left) Pattern Separation: When experiencing events of a similar nature (e.g., receiving a rose versus a different flower, accompanied by similar music), the hippocampus transforms overlapping cortical representations into distinct neural codes. Information is transmitted from the entorhinal cortex (EC) through the dentate gyrus (DG), where it undergoes sparse coding to orthogonalize similar inputs, thereby reducing interference. This dissociated representation is stored in CA3 and subsequently routed through CA1, thereby creating distinct memory indices (Index A and Index B) for experiences that are similar but not identical. This process enables the brain to differentiate between highly similar episodes and prevent memory interference. (Right) Pattern Completion: During the process of memory retrieval, the presentation of a partial cue (e.g., the sighting of a single rose) has been demonstrated to reaktivate the complete memory trace. The partial input enters through EC and activates a subset of the original pattern in CA3. The extensive recurrent collateral connections of CA3 form an attractor network, and this partial activation triggers pattern completion, thereby reconstructing the full memory representation. CA1 integrates this completed hippocampal representation with direct cortical input, thereby enabling the retrieval of the complete episodic memory, including contextual details and associated experiences. This mechanism enables the recollection of memories from fragmentary or degraded cues, thereby supporting robust retrieval even in instances where sensory input is incomplete. Collectively, these complementary processes—pattern separation during encoding and pattern completion during retrieval—enable the hippocampus to both store similar memories without interference and reconstruct complete memories from partial information. This forms the computational foundation of episodic memory systems. Bakker et al. (2008) detected signals consistent with pattern separation in the DG-CA3 region in humans using high-resolution fMRI. This finding lends support to theoretical models suggesting that different subregions of the hippocampal circuit share these two complementary functions (separation and completion). Indeed, this functional differentiation within the hippocampal circuit has been demonstrated in several studies. For instance, Sun et al. (2017) observed that the DG-CA3 pair exhibited pattern separation signals, while the CA3-CA1 and CA1-subiculum connections exhibited pattern completion signals. In consideration of these results, it can be posited that the hippocampal subregions concurrently engage in both pattern separation and pattern completion functions. A recent study has provided substantial support for this hypothesis. The present study sought to examine these functions of the hippocampus in greater detail by leveraging intracranial electroencephalography (EEG) data. The study’s findings indicated that the anterior hippocampus exhibits heightened activity during the initial stages of information extraction, while the posterior hippocampus plays a pivotal role in maintaining these representations in subsequent stages (To et al., 2025). Capture of Memory Through Engrams The foundation of index theory has been further solidified by the elaboration of engram cells (Guskjolen and Cembrowski, 2023; Josselyn and Tonegawa, 2020; Kitamura et al., 2017; Lei et al., 2025; Lopez et al., 2024; Miry et al., 2021; Tanaka and McHugh, 2018; Tanaka et al., 2014). Specifically, the advent of technologies such as optogenetic manipulation has demonstrated that particular groups of neurons in the hippocampus are capable of encoding specific memories (Liu et al., 2012; Ramirez et al., 2013; Redondo et al., 2014). Engrams are defined as specific memories or memories in general. The re-excitation of engrams has been demonstrated to result in the reactivation of the neocortical network (Gershman et al., 2025). These findings suggest that engrams function as an effective index for episodic memories and that the hippocampus plays a pivotal role in memory retrieval. One of the most concrete pieces of evidence in this field comes from the experimental work of Kolibius and colleagues (2023), who identified ”Episode-Specific Neurons” (ESNs) in the human medial temporal lobe. These ESNs encompass all the contextual components of a memory. These characteristics distinguish them from classical ”concept neurons,” which respond to a single concept independently of context (Kolibius et al., 2023). However, this is not a static distinction; a theoretical proposal put forward in a subsequent compilation published by the same team (2025) postulates that these two neuron types are in a dynamic relationship. According to this model, ESNs undergo a transformation into concept neurons over time. This transformation occurs through the repetition of overlapping experiences. That is to say, an ESN that initially encodes a specific memory gradually loses its selectivity for the details unique to that memory and generalizes to the fundamental concept common to many memories. This mechanism provides a direct neural model of how an episodic trace transforms into semantic information, suggesting that the hippocampus may be a dynamic processor capable of generating abstract information from experiences (Kolibius et al., 2025). Specifically, it can be posited that responses specific to individual events undergo a shift over time, moving toward the common features of conceptual clusters. This process of integration occurs within the same architecture, merging episodic traces with semantic generalization. This approach has been supported by computational models in recent years (Spens and Burgess, 2024). This framework is consistent with the hypothesis that ”key-like” cues selectively activate cortical ”value” representations during episodic recall. Indeed, neurons demonstrating high selectivity for concept and spatial information have been observed to co-occur in the medial temporal lobe. This finding indicates that conceptual abstraction can adaptively integrate with contextual dimensions (Mackay et al., 2024). These results may be interpreted as providing support for the direction of generalization predicted by Kolibius. The existing literature on single-neuron studies of memory encoding in the human medial temporal lobe (MTL) presents a complex and seemingly contradictory picture regarding fundamental processes such as hippocampal indexing and pattern discrimination. In their seminal work, Rey et al. (2023) profoundly challenged the long-standing assumption of context-specific pattern discrimination at the single-neuron level. They demonstrated that representations of a concept (e.g., a specific person) in MTL neurons are almost entirely independent of narrative context and that conjunctive coding is almost entirely absent. This finding suggests that the fundamental components of memory are immutable ”concept cells” that can be flexibly reassembled (Rey et al., 2025). Conversely, Cao et al. elucidate a preceding stage in the formation of these abstract representations, demonstrating that an object’s memorability is contingent upon a ”region-based feature code” that is informed by its abstract visual properties, thereby establishing a nexus between perceptual processing and memory formation. The integration of these two findings unveils a hierarchical model in which the MTL initially executes feature-based filtering (Cao et al., 2025) and subsequently converts this into an invariant semantic representation (Rey et al., 2025), which serves as the foundation for generalization. This framework prompts the following question: How does the hippocampal index link an episode by assembling these pre-existing, invariant conceptual building blocks and encoding the relationships between them, rather than creating entirely new and context-specific representations for each experience? These findings collectively demonstrate an evolution of the hippocampal index theory from a static model to a temporally and spatially reorganizable information retrieval system. When ESN-concept conversion, key-value architecture, and DG–CA3 dynamics are evaluated in concert, they yield the conclusion that the hippocampus not only records past experiences but also produces new conceptual structures and cognitive flexibility from these experiences. This communication between DG and CA3 governs a critical decision of the memory system: to encode a new memory or to retrieve an existing one. In instances where the input is familiar and partial, the auto-associative network in CA3 engages in favor of completion. Conversely, when the input is new or ambiguous, the sparse and orthogonal representations produced by the DG promote the formation of a new index. The methodology underlying the creation of the new index has recently become a topic of more concrete discussion, facilitated by the results of a recent study. The present study on avian foraging behavior revealed that specific, transient ”barcode-like” firing patterns emerge in the hippocampus of these birds during food-caching events (Chettih et al., 2024). These ”barcodes” function as a unique identifier for each memory, and they are reactivated instantaneously when the associated storage location is recalled. Of particular significance is the finding that these barcodes appear to be independent of the activity of place cells, which encode location. This phenomenon persists even when comparing two storage memories of a remarkably similar nature, despite their evident disparities in ”barcode” production. The finding that the hippocampus assigns a unique neural identity to each episodic memory, similar to a computer’s hash code, is quite striking because it can be considered direct observational evidence for this index theory. While these molecular-level studies have elucidated numerous hitherto obscure aspects of the recall system, several significant elements remain to be elucidated. Notably, there is a paucity of research addressing critical questions surrounding the selection and formation of indexes, as well as the mathematical underpinnings of pattern completion. In the context of multidisciplinary fields, such as neuroscience, computational approaches and models have emerged as pivotal research instruments. The advent of machine learning and artificial intelligence has profoundly impacted various facets of human existence, rendering them indispensable components of contemporary living. This progression has been fueled by the integration of biological systems’ knowledge into machine engineering, resulting in the development of highly efficient, rapidly learning systems that deliver precise results. Concurrently, technological advancements have enabled the development of machines capable of learning and retaining information, emulating the cognitive processes observed in humans. This approach enables the acquisition of novel perspectives on inquiries that have yet to be elucidated within the domain of biological systems. In the subsequent sections of this article, we explore the potential for refining and diversifying our inquiry by adopting an alternative approach to the complex processes in question. This involves the utilization of computational methodologies, with a particular emphasis on machine learning techniques, in conjunction with the most recent biological evidence concerning the recall process. Additionally, we offer a synopsis of insightful perspectives on the manner in which these issues are addressed in contemporary machine and artificial intelligence systems, as well as in artificial learning models that draw inspiration from neuroscience and incorporate memory mechanisms. Can Machines Remember Like Humans? The conceptual underpinnings of pattern completion in computational neuroscience were initially established by Hopfield’s (1982) seminal attractor network model. This model was inspired by the recurrent connections of the CA3 region and was developed by researchers such as Treves and Rolls (1991). It incorporates more biologically realistic mechanisms, such as sparse coding, to enhance storage capacity (Rolls and Treves, 1991; Treves and Rolls, 1994). However, fundamental limitations of these classical networks, such as low storage capacity and a tendency to generate spurious attractors that reduce reliability as the number of stored patterns increases, have led researchers in the field to develop more advanced and efficient memory architectures. This research led to the development of modern Hopfield networks, which exponentially increase storage capacity. However, these new models also posed a biological dilemma: the mathematical formulation of these networks required many-body synaptic connections that appeared to be biologically implausible. Krotov and Hopfield (2020) developed a ”microscopic theory” that resolves this contradiction. Their work demonstrated that the same exponential capacity can be achieved using only standard two-body synaptic connections by employing an additional hidden layer of neurons (Krotov & Hopfield, 2020). The conceptual parallelism proposed by this model for the hippocampal circuits (entorhinal cortex → CA3/CA1) establishes a bridge between abstract computational theories and the brain’s physical architecture, thereby creating new avenues for research. Deep Network Models Contemporary methodologies in the field of memory research endeavor to illuminate the deficiencies in recall and pattern completion processes by employing computational models that emulate the distinctive characteristics of hippocampal circuits. For instance, Kanagamani et al. (2023) propose a novel methodology for modeling pattern separation and completion processes. The deep neural network they developed consists of two primary modules: an autoencoder representing cortico-hippocampal projections and a dynamic recall module that calculates the ”familiarity” level of a stimulus. This architecture effectively emulates the functions of the DG and CA3 regions, which play a pivotal role in recall processes. The autoencoder is employed to simulate pattern separation by reducing high-dimensional cortical input to a compressed hidden layer. A hill-climbing mechanism is then utilized to reconstruct a complete memory from a noisy cue and simulate pattern completion. Consequently, they were able to successfully reconstruct the original clean image from the noisy inputs. In addition, a particularly salient contribution of the model under consideration is its capacity to simulate Alzheimer’s disease (AD) pathology. The researchers simulated AD conditions by damaging a specific percentage of neurons in the network’s encoder layer. While the model demonstrated its capacity to recall a complete pattern under normal conditions and in the face of noisy inputs, it produced results under damaged conditions that strikingly matched clinical observations. Specifically, under conditions of moderate damage, the subject produced word errors, such as saying ”odd number” instead of ”nine.” This phenomenon has been frequently observed in patients with Alzheimer’s disease (AD), who often recall the superordinate category instead of a specific instance belonging to a concept. The model demonstrated an effective representation of severe cognitive impairment when the damage level was augmented. This was evidenced by the model’s tendency to produce a ”zero response,” which can be defined as an utterance such as ”I don’t know.” These findings demonstrate the model’s capacity to elucidate not only healthy memory processes but also specific cognitive impairments caused by neurodegenerative diseases at a mechanical level. Generative Models The examination of the theory that memory is a reconstructive process through computational models is a growing area of interest. Spens and Burgess (2024) endeavored to furnish a more mechanistic elucidation for the reconstructive nature of memory through the model they developed. The model is predicated on the hypothesis that replay signals from an auto-associative network representing the hippocampus train a generative network based on a variational autoencoder (VAE) representing the neocortex. In this process, the VAE learns a compressed latent representation that captures the statistical structure and essence (”schemas”) of experiences. The act of memory recall is conceptualized as the process of ”reconstructing” a sensory experience via a decoder based on this latent representation. The model’s simulations yield outcomes that align with the foundational tenets of systemic consolidation theory. As the replay process progresses, the generative network (neocortex) gradually assumes control of the recall function. This finding provides a compelling explanation for the phenomenon that lesions to the hippocampus, a brain region involved in memory formation, result in impaired new memory formation while largely preserving well-consolidated old memories. A seminal aspect of the model pertains to its elucidation of schema-based memory distortions, a crucial contribution to the field. The producer network learns the statistical average of experiences, and memories are ”pulled” toward this schema during recall. This mechanism inherently engenders memory errors of the ”boundary extension” variety, thereby elucidating the escalating prevalence of such distortions as consolidation progresses. Consequently, this model bridges important implications by transforming memory recall from a static ”playback” action into a dynamic ”reconstruction” process that shares the same generative mechanism with imagination and inference. This unifying framework suggests that while the hippocampus optimizes its limited storage capacity to encode novel and unpredictable sensory details, the neocortex constructs a statistical model of the world. Consequently, a shared computational foundation is established, integrating previously distinct cognitive domains such as memory, imagination, and planning. It can be posited that the process of hippocampal replay functions not only to fortify existing traces but also to utilize experiences as ”building blocks,” thereby facilitating the creation of novel combinations. Furthermore, the results indicate that during replay, new neural firing fields are placed in distant locations corresponding to newly discovered objects. This finding suggests a relationship between memory recall and the capacity to formulate novel scenes from existing components, thereby facilitating generalization to previously unfamiliar environments (Spens et al., 2024). Hybrid Architectures This generative approach is further advanced by the VAE+Modern Hopfield Network (VAE+MHN) model developed by Jun et al. (2024) to address catastrophic forgetting, one of the most fundamental challenges in machine learning. The hybrid system has been developed as an application of the biologically based Complementary Learning Systems (CLS) theory (McClelland et al., 1995). This theory posits that the brain possesses the capacity to retain prior knowledge while concurrently acquiring new information. A critical facet of the research pertains to the exploration of the phenomenon of memory or the failure to recall prior information, often referred to as ”catastrophic forgetting.” This aspect is approached from a novel perspective, offering a fresh approach to understanding this complex issue. This model is designed to manage two fundamental functions by dividing them into distinct modules. The Modern Hopfield Network (MHN) is employed to emulate the function of the hippocampus in rapid learning and pattern recognition, while the Variational Autoencoder (VAE) models the neocortex’s capacity for slow, abstract learning and partial pattern completion. The researchers evaluated the effectiveness of this architecture on Split-MNIST, a standard benchmark task in the field of continuous learning. In this task, the model must gradually learn new digit classes without forgetting the previous ones. The findings indicate that the model exhibits a high degree of efficacy in mitigating forgetting, with near-perfect accuracy. Of particular significance is the empirical corroboration of this functional separation through representational analyses. While the internal representations of the MHN module guide pattern separation, the VAE module ensures pattern completion. This work is noteworthy in that CLS theory is not only an explanatory framework but also that a functional architecture can be created to solve questions related to the concept of ”plasticity” in machine learning. Another significant approach that offers a more fundamental imitation of hippocampal circuits is the HiCL (Hippocampal-Inspired Continual Learning) architecture developed by Kapoor et al. (2025). The model under consideration is predicated on a DG-Gated Mixture-of-Experts structure that directly mimics the brain’s trisynaptic circuit. At the core of the architecture is a Dentate Gyrus (DG)-like layer that highly orthogonalizes the inputs, thereby performing the pattern separation function. Subsequently, a CA3-like module, implemented as a two-layer MLP autoencoder, undertakes the pattern completion task with an attractor-like reconstruction. In conclusion, a CA1-like module integrates this completed pattern with cortical inputs, functioning as the integration and consolidation stage. A notable innovation of the model is its ability to function without a dedicated network for task guidance, a feature that sets it apart from previous models. Conversely, the sparse representations generated by the DG layer dynamically direct inputs to the pertinent ”expert” network, contingent on task prototypes that have been learned. To address the challenge of catastrophic forgetting, the architecture incorporates built-in reinforcement mechanisms, including prioritized replay and Elastic Weight Consolidation (EWC). This holistic approach offers a compelling illustration of how the stability-plasticity dilemma inherent in biological systems’ continuous learning processes can be effectively translated into a practical engineering solution. Indeed, this architecture achieves competitive results on standard continuous learning tasks, such as Split CIFAR-10, while accomplishing this with lower computational cost. This outcome is significant in demonstrating how important outcomes can be achieved when biological principles are synthesized with engineering efficiency. Refer to Figure [2]. This figure explores the correlation between architectural correspondence in representation learning models and biological complementary learning systems (CLS). The modern representation learning framework (Left) illustrates the computational pipeline, which begins with input and progresses through hierarchical encoding (e.g., CNNs, Vision Transformers, and VAE encoders) and culminates in compressed latent representations. From this latent space, two parallel mechanisms operate: pattern completion (via Modern Hopfield Networks, autoencoders, or graph traversal) performs associative recall and reconstructs complete patterns from partial cues, while pattern separation (via sparse layers or DG-like modules) orthogonalizes representations to reduce interference. The consolidation of memory is achieved through the integration of these processes by replay mechanisms and gradient-based updates, facilitating the transition from episodic to semantic knowledge while mitigating the risk of catastrophic forgetting. Decoding involves the reconstruction of outputs through the utilization of generative models, accompanied by feedback loops (illustrated by gray arrows) that facilitate the refinement of earlier encoding stages by consolidated memories. The following illustration (Right) depicts a biological complementary learning systems architecture, illustrating the neural implementation of analogous processes. Neocortical regions (PRC, PHC) are responsible for hierarchical feature processing, with the entorhinal cortex (EC) acting as a conduit between the cortex and the hippocampus. The hippocampus implements pattern separation in the dentate gyrus (DG) through sparse coding, pattern completion in CA3 via recurrent attractor dynamics, and integration in CA1. The process of memory consolidation is facilitated by the replay mechanism in the hippocampus, which systematically transfers information back to the neocortex, as illustrated by the orange feedback arrows. This process enables the consolidation of systems, thereby transforming episodic memories into semantic knowledge. The subiculum (SBC) is responsible for channeling processed information to output pathways. The application of brown brackets serves to underscore pivotal correspondences, including the following: encoding and neocortical processing, latent space and entorhinal gateway, pattern completion and CA3 attractors, pattern separation and DG sparse coding, and consolidation and hippocampal replay. Both systems balance fast hippocampal-like learning (rapid, instance-specific) with slow neocortical-like learning (gradual, statistical) to achieve continual learning without catastrophic forgetting. The presence of bidirectional arrows in the latent/pattern processing stages is indicative of dynamic interactions that facilitate flexible memory operations in both biological and artificial intelligence systems. Separated Representations and Abstract Cognitive Maps The Tolman-Eichenbaum Machine (TEM), developed by Whittington and colleagues (2020), examines these approaches at a more abstract computational level, unifying the two long-studied core functions of the hippocampus—spatial navigation and relational memory—under a single framework. The model is predicated on the notion that experiences can be decomposed into two distinct factors: sensory content (”what”) and the structural relationships between these contents (”where” or ”how”). This dissociation is predicated on the hypothesis that the medial entorhinal cortex encodes structural information (e.g., the geometry of a space or a conceptual graph), while the cells of the hippocampus link this structural foundation to specific sensory representations. In essence, this process enables the system to extrapolate the structural information it has learned from one context (e.g., the layout of a maze or the hierarchy of a family tree) to another situation that is entirely devoid of sensory data. Consequently, the model possesses the capacity to formulate complex inferences based on observations, even in previously unencountered environments. To illustrate, it can instantaneously predict all remaining relationships in a novel social setting by observing a single relationship. Researchers have presented a theoretical framework with the term ”temporal-event model” (TEM), which they use to predict the emergence of various neural representations in the entorhinal cortex. These representations include grid, boundary, and object-vector cells. This model elucidates the underlying mechanism that facilitates the brain’s navigation of physical spaces and the establishment of conceptual relationships on an abstract ”cognitive map,” thereby offering a fundamental principle regarding the compositional nature of memory. A recent illustration of this approach is the HippoRAG framework developed by Gutierrez and colleagues (2024) (Jiménez Gutiérrez et al., 2024). This system directly translates hippocampal indexing theory into a practical engineering solution for large language models (LLMs), synergistically mimicking the roles of brain regions in continuous language switching (CLS) theory. In the operation of the framework, the initial step involves the modeling of the neocortex’s role in processing perceptual input. To this end, an LLM converts raw text documents into a structured knowledge graph composed of discrete concepts and relationships. This graph serves as an artificial ”hippocampal index.” During the recall phase, the system implements a Personalized PageRank (PPR) algorithm on this knowledge graph, leveraging key concepts derived from the user’s query. This mechanism is analogous to the hippocampus’s efficient search through neural pathways to activate related memories, thereby performing pattern completion. Evaluations on challenging tasks requiring the integration of different pieces of information, such as multi-hop question-answering (QA), have shown HippoRAG to outperform current state-of-the-art methods. Moreover, its lower cost and faster speed compared to single-step retrieval and iterative methods also demonstrate the efficiency of this biologically inspired design. HippoRAG exemplifies the utilization of theoretical models such as CLS as a tangible engineering framework, thereby augmenting the knowledge-based and context-aware reasoning capabilities of contemporary artificial intelligence systems. This augmentation is achieved with enhanced precision and computational efficiency, addressing a significant challenge in the field. The subsequent stage in the constructive memory understanding process entails the design of architectures capable of modeling continuous and multi-experience flows that extend beyond static datasets. In this regard, seminal contributions such as HippoMM (Zhao et al., 2025) seek to address this challenge by directly translating the principles of the hippocampus into a computational framework. HippoMM accomplishes this objective by integrating three fundamental innovations. The first of these mechanisms draws inspiration from the hippocampus’s pattern separation ability, providing a dynamic mechanism that divides continuous visual-auditory streams into meaningful events. Secondly, it emulates a biological consolidation process that models the transition from short-term to long-term memory. In this process, it utilizes a method called ”semantic replay,” which converts detailed perceptual traces (short-term memory) into abstract semantic summaries (long-term memory) via a Large Language Model (LLM). Finally, akin to the human associative recall process, it employs a cross-modal retrieval mechanism that facilitates the retrieval of related information in one modality (e.g., a sound) from a cue in another modality (e.g., a related image). This approach signifies a substantial shift in the concept of artificial intelligence memory, moving it from a static information storage task toward the goal of becoming a dynamic and integrated internal model.

Synthesis and Future Challenges As demonstrated, a range of computational models, including deep autoencoders, generative networks, hybrid CLS systems, and compositional replay approaches, offer multidimensional perspectives on unsolved problems in memory research. These studies enhance the memory efficiency of machine learning systems and facilitate a more profound comprehension of the role of hippocampal mechanisms in time, structural integration, and recall processes. However, as Rolls (Rolls and Treves, 2024) emphasizes, this convergence also highlights fundamental computational differences between local and associative learning rules in the brain and backpropagation-based approaches in contemporary artificial intelligence. This scenario underscores a fundamental challenge known as the ”modeler’s dilemma”: How should the balance between biological realism and computational efficiency be achieved? The increasing intricacy of these models is surpassing the capabilities of conventional computer architectures, underscoring the necessity for novel computational paradigms, such as neuromorphic hardware (Casanueva-Morato et al., 2023), to effectively execute such brain-like algorithms in the future. There are some important questions about the reconstructive nature of memory that remain unanswered and need to be clarified. 1-How exactly are indexes formed, and can the decision-making mechanism behind this formation be clearly explained? 2-Should indexes and episode-specific neurons be considered as two separate elements of a larger complementary system? 3-How can advanced artificial intelligence algorithms help us simulate and better understand the nature of reconstructive memory? Consequently, our endeavor to comprehend memory and our endeavors to artificially create it should not be regarded as discrete entities. While the biological foundations offered by neuroscience are pushing the boundaries of artificial intelligence, the models produced by artificial intelligence are also increasing the quality and detail of the questions we ask about memory, one of the most fundamental building blocks of consciousness. The future of illuminating the dark parts of memory lies in the synthesis of these two disciplines.

Bakker, A., Kirwan, C. B., Miller, M., & Stark, C. E. (2008). Pattern separation in the human hippocampal CA3 and dentate gyrus. Science, 319 (5870), 1640–1642. https://doi.org/10.1126/science.1152882 Cao, R., Brunner, P., Chakravarthula, P.N. et al. A neuronal code for object representation and memory in the human amygdala and hippocampus. Nat Commun 16, 1510 (2025). https://doi.org/10.1038/s41467-025-56793-y Casanueva-Morato D., A. Ayuso-Martinez, J. P. Dominguez-Morales, A. Jimenez-Fernandez and G. Jimenez-Moreno, ”Spike-based computational models of bio-inspired memories in the hippocampal CA3 region on SpiNNaker,” 2022 International Joint Conference on Neural Networks (IJCNN), Padua, Italy, 2022, pp. 1-9, doi: 10.1109/IJCNN55064.2022.9892606. Chettih, S. N., Mackevicius, E. L., Hale, S., & Aronov, D. (2024). Barcoding of episodic memories in the hippocampus of a food-caching bird. Cell, 187 (8), 1922–1935.e20. https://doi.org/10.1016/j.cell.2024.02.032 Gershman, S. J., Fiete, I., & Irie, K. (2025). Key-value memory in the brain . Neuron, 113(11), 1694–1707. https://doi.org/10.1016/j.neuron.2025.02.029 Guskjolen, A., & Cembrowski, M. S. (2023). Engram neurons: Encoding, consolidation, retrieval, and forgetting of memory. Molecular Psychiatry, 28 (8), 3207–3219. https://doi.org/10.1038/s41380-023-02137-5 Guzman, S. J., Schlögl, A., & Jonas, P. (2016). Synaptic mechanisms of pattern completion in the hippocampal CA3 network. Science, 353 (6304), 1117–1123. https://doi.org/10.1126/science.aaf1836 Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79 (8), 2554-2558. https://doi.org/10.1073/pnas.79.8.2554 Hunsaker, M. R., Kesner, R. P. (2013). The operation of pattern separation and pattern completion processes associated with different attributes or domains of memory . Neuroscience and biobehavioral reviews. https://doi.org/10.1016/j.neubiorev.2012.09.014 Jiménez Gutiérrez, B. M., Guo, T., & Schwartz, J. (2024). HippoRAG: Neurobiologically inspired long-term memory for large language models . arXiv. https://doi.org/10.48550/arXiv.2405.14831 Josselyn, S. A., & Tonegawa, S. (2020). Memory engrams: Recalling the past and imagining the future. Science, 367 (6473), eaaw4325. https://doi.org/10.1126/science.aaw4325 Jun, H., Lee, J., Jang, G., & Hwang, S. J. (2024). A neural network model of complementary learning systems: Pattern separation and completion for continual learning . arXiv. https://doi.org/10.48550/arXiv.2507.11393 Kanagamani, M. D., Jang, J., Lee, S., & Lee, S. W. (2023). A deep network-based model of hippocampal memory functions under normal and Alzheimer’s disease conditions. Frontiers in Neural Circuits, 17, 1092933. https://doi.org/10.3389/fncir.2023.1092933 Kapoor, K., Mackey, W., Aloimonos, Y., & Lin, X. (2025). HiCL: Hippocampal-Inspired Continual Learning (arXiv preprint arXiv:2508.16651). https://doi.org/10.48550/arXiv.2508.16651 Kitamura, T., Ogawa, S. K., Roy, D. S., Okuyama, T., Morrissey, M. D., Smith, L. M., & Tonegawa, S. (2017). Engrams and circuits crucial for systems consolidation of a memory. Science, 356 (6333), 73–78. https://doi.org/10.1126/science.aam6808 Krotov, D. & Hopfield, J.~J. (2020). Large associative memory problem in neurobiology and machine learning (arXiv preprint arXiv:2008.06996). https://doi.org/10.48550/arXiv.2008.06996 Kolibius, L. D., Roux, F., Parish, G., Ter Wal, M., Van Der Plas, M., Chelvarajah, R., Sawlani, V., Rollings, D. T., Lang, J. D., Gollwitzer, S., Walther, K., Hopfengärtner, R., Kreiselmeyer, G., Hamer, H., Staresina, B. P., Wimber, M., Bowman, H., & Hanslmayr, S. (2023). Hippocampal neurons code individual episodic memories in humans. Nature human behaviour, 7 (11), 1968–1979. https://doi.org/10.1038/s41562-023-01706-6 Kolibius, L. D., Josselyn, S. A., & Hanslmayr, S. (2025). On the origin of memory neurons in the human hippocampus. Trends in Cognitive Sciences, 29(5), 421-433. https://doi.org/10.1016/j.tics.2025.01.013 Lei, B., Kang, B., Hao, Y., Yang, H., Zhong, Z., Zhai, Z., & Zhong, Y. (2024). Reconstructing a new hippocampal engram for systems reconsolidation and remote memory updating. Neuron, 113(3), 471–485. https://doi.org/10.1016/j.neuron.2024.11.010 Leutgeb, J. K., Leutgeb, S., Moser, M. B., & Moser, E. I. (2007). Pattern separation in the dentate gyrus and CA3 of the hippocampus. Science, 315 (5814), 961–966. https://doi.org/10.1126/science.1135801 Liu, X., Ramirez, S., Pang, P. T., Puryear, C. B., Govindarajan, A., Deisseroth, K., & Tonegawa, S. (2012). Optogenetic stimulation of a hippocampal engram activates fear memory recall. Nature, 484 (7394), 381–385. https://doi.org/10.1038/nature11028 Lopez, M. R., Wasberg, S. M. H., Gagliardi, C. M., Normandin, M. E., & Muzzio, I. A. (2024). Mystery of the memory engram: History, current knowledge, and unanswered questions. Neuroscience & Biobehavioral Reviews, 159, 105574. https://doi.org/10.1016/j.neubiorev.2024.105574 Mackay, S., Reber, T. P., Bausch, M., Boström, J., Elger, C. E., & Mormann, F. (2024). Concept and location neurons in the human brain provide the ‘what’ and ‘where’ in memory formation. Nature Communications, 15, 7926. https://doi.org/10.1038/s41467-024-52295-5 Marr, D. (1971). Simple memory: A theory for archicortex. Philosophical Transactions of the Royal Society of London. B, Biological Sciences, 262 (841), 23–81. https://doi.org/10.1098/rstb.1971.0078 McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. Psychological Review, 102 (3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419 Miry, O., Li, J., & Chen, L. (2021). The quest for the hippocampal memory engram: From theories to experimental evidence. Frontiers in Behavioral Neuroscience, 14, 632019. https://doi.org/10.3389/fnbeh.2020.632019 Ramirez, S., Liu, X., Lin, P. A., Suh, J., Pignatelli, M., Redondo, R. L., Ryan, T. J., & Tonegawa, S. (2013). Creating a false memory in the hippocampus. Science (New York, N.Y.), 341 (6144), 387–391. https://doi.org/10.1126/science.1239073 Redondo, R. L., Kim, J., Arons, A. L., Ramirez, S., Liu, X., & Tonegawa, S. (2014). Bidirectional switch of the valence of a memory engram. Nature, 513 (7518), 426-430. https://doi.org/10.1038/nature13725 Rey, H. G., Panagiotaropoulos, T. I., Gutierrez, L., Chaure, F. J., Nasimbera, A., Cordisco, S., Nishida, F., Valentin, A., Alarcon, G., Richardson, M. P., Kochen, S., & Quian Quiroga, R. (2025). Lack of context modulation in human single neuron responses in the medial temporal lobe. Cell reports, 44 (1), 115218. https://doi.org/10.1016/j.celrep.2024.115218 Rolls, E. T. (2013). The mechanisms for pattern completion and pattern separation in the hippocampus. Frontiers in Systems Neuroscience, 7, 74. https://doi.org/10.3389/fnsys.2013.00074 Rolls, E. T., & Treves, A. (1991). The storage of memory in the hippocampal network. In J. L. Davis & H. Eichenbaum (Eds.), Olfaction: A model system for computational neuroscience (pp. 283–300). MIT Press. ISBN 978-0262041249 Rolls, E. T., & Treves, A. (2024). A theory of hippocampal function: New developments. Progress in Neurobiology, 238, 102636. https://doi.org/10.1016/j.pneurobio.2024.102636 Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: Remembering the past and imagining the future. Philosophical Transactions of the Royal Society B: Biological Sciences, 362 (1481), 773-786. https://doi.org/10.1098/rstb.2007.2087 Spens, E., & Burgess, N. (2024). A generative model of memory construction and consolidation. Nature human behaviour, 8 (3), 526–543. https://doi.org/10.1038/s41562-023-01799-z Sun, Q., Sotayo, A., Cazzulino, A. S., Snyder, A. M., Denny, C. A., & Siegelbaum, S. A. (2017). Proximodistal heterogeneity of hippocampal CA3 pyramidal neuron intrinsic properties, connectivity, and reactivation during memory recall. Neuron, 95 (3), 656–672.e5. https://doi.org/10.1016/j.neuron.2017.07.012 Tanaka, K. Z., & McHugh, T. J. (2018). The hippocampal engram as a memory index. Journal of Experimental Neuroscience, 12, 1179069518815942. https://doi.org/10.1177/1179069518815942 Tanaka, K. Z., Pevzner, A., Hamidi, A. B., Nakazawa, Y., Graham, J., & Wiltgen, B. J. (2014). Cortical representations are reinstated by the hippocampus during memory retrieval. Neuron, 84 (2), 347–354. https://doi.org/10.1016/j.neuron.2014.09.037 Teyler, T. J., & DiScenna, P. (1986). The hippocampal memory indexing theory. Behavioral Neuroscience, 100 (2), 147–154. https://doi.org/10.1037/0735-7044.100.2.147 Teyler, T. J., & Rudy, J. W. (2007). The hippocampal indexing theory and episodic memory: Updating the index. Hippocampus, 17 (12), 1158–1169. https://doi.org/10.1002/hipo.20350 To, T. V., Wang, D. X., Wolfe, C. B., & Lega, B. C. (2025). Neurophysiological evidence of human hippocampal longitudinal differentiation in associative memory. Nature communications, 16 (1), 6845. https://doi.org/10.1038/s41467-025-61464-z Treves, A., & Rolls, E. T. (1994). Computational analysis of the role of the hippocampus in memory. Hippocampus, 4 (3), 374–391. https://doi.org/10.1002/hipo.450040319 Whittington, J. C. R., Muller, T. H., Mark, S., Chen, G., Barry, C., Burgess, N., & Behrens, T. E. J. (2020). The Tolman-Eichenbaum machine: Unifying space and relational memory through generalization in the hippocampal formation. Cell, 183 (5), 1249–1263.e23. https://doi.org/10.1016/j.cell.2020.10.024 Yassa, M. A., & Stark, C. E. (2011). Pattern separation in the hippocampus. Trends in Neurosciences, 34 (10), 515–525. https://doi.org/10.1016/j.tins.2011.06.006 Yassa, M. A., Lacy, J. W., Stark, S. M., Albert, M. S., Gallagher, M., & Stark, C. E. (2011). Pattern separation deficits associated with increased hippocampal CA3 and dentate gyrus activity in nondemented older adults. Hippocampus, 21 (9), 968–979. https://doi.org/10.1002/hipo.20808 Zhao, R., Liu, B., Yu, H., Singh, C., & Yu, Y. (2025). HippoMM: Hippocampal-inspired multimodal memory for long audiovisual event understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . https://doi.org/10.48550/arXiv.2504.10739 Information & Authors Information Version history Copyright This work is licensed under a Non Exclusive No Reuse License.

Authors Metrics & Citations Metrics Article Usage 783views 171downloads Citations Download citation Kaan Cengiz, Guldal Gulec Suyen. Reassessment of the Hippocampal Index Theory: A Synopsis of the Present Evidence and Computational Models. Authorea. 29 November 2025. DOI: https://doi.org/10.22541/au.176442317.74625166/v1 DOI: https://doi.org/10.22541/au.176442317.74625166/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu.