External Hippocampus Framework

• External Hippocampus Framework is a computational model defining the hippocampus as an external, rapidly updating memory system interfacing with neocortical modules.
• It employs data-structural hierarchies, one-shot learning, and topological map formation to simulate episodic encoding, sequence storage, and cognitive schema integration.
• The framework has diverse applications, informing AI reasoning, neuromorphic hardware design, and clinical neuroimaging for memory-related disorders.

The External Hippocampus Framework encompasses algorithmic, computational, and topological models that formalize the hippocampus as an external memory system interfacing with neocortical and sensorimotor modules. This framework has been operationalized across biological, neuromorphic, machine learning, robotics, cognitive modeling, and neuroimaging domains. Its core principle is the separation of rapidly-encoded, interference-prone episodic memory (hippocampus) from slowly-formed, robust semantic or perceptual invariants (neocortex), instantiated through specialized data structures, topological constructions, attractor dynamics, and structural morphometrics. Notable implementations include data-structural hierarchies rooted in Hubel–Wiesel modules, online one-shot sequence storage, topological cognitive map analysis, LLM reasoning guidance, bio-inspired memory hardware, vision-language episodic memory architectures, and anatomically grounded morphometric coordinate frameworks.

1. Algorithmic and Data Structural Foundations

The original formalism of the External Hippocampus Framework as presented by Leibo et al. models the cortex–hippocampus axis as a cascade of Hubel–Wiesel Modules (HWMs) (Leibo et al., 2015). Each layer receives a $d$-dimensional input vector, pools over banks of HW modules via similarity functions (normalized dot-product), and produces signatures $\mu^{(\ell)}(x) \in \mathbb{R}^{K_\ell}$. At each level, HWMs function as data structures supporting two operations:

• INSERT: adding new templates to a class-specific set $\mathcal{T}_k$;
• QUERY: evaluating a pooling function over similarity scores between a query $x$ and stored templates in $\mathcal{T}_k$.

Critical to the framework is the time-scale separation in memory encoding:

• Cortical learning (HWMs up to IT/PRC) is approximated by slow, iterative PCA (Oja’s rule), resulting in compressed, invariant representations that generalize across stimulus variability.
• Hippocampal learning (final HWM layer) is realized via extremely rapid one-shot insertion, employing either random projection (Johnson–Lindenstrauss lemma) or locality sensitive hashing (LSH), enabling the storage of high-fidelity episodic episodes with immediate queryability and capacity scaling as $N \lesssim \exp(cs\epsilon^2)$ for projections of dimension $s$.

This constructs the hippocampus as an "external" data-structure, receiving deeply processed but information-rich signatures from the neocortical hierarchy and providing rapid, interference-limited insert/query associative memory operations (Leibo et al., 2015).

2. Sequence Storage, Pattern Separation, and Online Forgetting

The "external hippocampus" instantiation in CRISP-theory derived models delivers an architecture for continuous one-shot online sequence storage (Melchior et al., 2019). Here, a sequence of patterns is hetero-associated through the following subregions:

• Sensory Input to Entorhinal Cortex (EC): non-plastic encoding of raw patterns into high-dimensional sparse codes.
• Dentate Gyrus (optional, for pattern separation): compresses temporal correlations, enabling robust encoding of correlated episodes by mapping into a higher-dimensional, sparse regime.
• CA3 (intrinsic, fixed recurrent "sequence machine"): serves as a cyclic sequence generator for target states, with online hetero-association between EC and CA3 completed by a one-shot Hebbian-descent rule:

  $\delta w_{ij} = -\eta (x_i - \mu_i) (h_j - t_j)$

with automatic asymptotic forgetting resulting as older memory traces are superseded by new updates, producing a power-law decay in retrieval quality.

Empirical sequence memory capacity is approximately $40\%$ of the number of CA3 neurons, and "dreaming"—offline, self-supervised sequence replay using noisy reconstructions—further decorrelates and stabilizes the EC–CA3 mappings without fresh external input (Melchior et al., 2019).

3. Topological Map Formation and Cognitive Schemas

The framework extends to spatial and cognitive mapping via topological data analysis and schema theory, establishing that the hippocampus encodes robust invariants at multiple scales (Dabaghian, 2019, Babichev et al., 2015). Here, the coactivity of place cells is mapped onto evolving simplicial complexes:

• Nerve (Čech) Complex: abstracts spatial environment as intersections of place fields, with Betti numbers corresponding to environment homology.
• Temporal Coactivity Complex: built from spike-train timing, yielding a persistent homology barcode whose stabilization time $T_{\min}$ quantifies spatial learning rate.

Four core schemas underpin the computational mapping:

1. Graph schema: binary connections for cell pairs with coactivity within a temporal window.
2. Simplicial schema: higher-order coactivity, enabling extraction of topological features (connectedness, loops).
3. Mereological schema: temporal coverage relations (containment).
4. RCC5 calculus: region connection logic with categorical spatial relations and motif detection.

Large-scale invariants (schema "integrals") such as Betti numbers, entropy metrics, and junction motifs, emerge rapidly—often before the full network of readout neurons is trained—allowing early cognitive map inference. Architecturally, these modules operate as independent subscribers on a shared spike-event stream, publishing map integrals to downstream planners (Dabaghian, 2019, Babichev et al., 2015).

4. External Hippocampus in AI and Neuromorphic Systems

Bio-inspired and neuromorphic systems implement the "external hippocampus" as an associative memory capable of rapid storage, cued recall, and intrinsic forgetting (Casanueva-Morato et al., 2022). On hardware such as SpiNNaker, spiking neural network modules map anatomical subregions:

• DG: binary cue decoder.
• CA3: split into cue and content populations, connected via STDP synapses implementing pairwise timing-dependent plasticity.
• CA1/Cortical interface: enables bidirectional recoding.

The design achieves sub-μJ energy per access, rapid (7–12 ms) one-shot learning and recall, efficient storage of both orthogonal and non-orthogonal patterns, and rehearsal-based forgetting mirroring biological processes. Limitations include capacity bottlenecks, lack of content-based associative recall, and absent theta/gamma oscillations (Casanueva-Morato et al., 2022).

In cognitive AI, the External Hippocampus framework also guides LLM reasoning by building topological cognitive maps of semantic state-space (Yan, 20 Dec 2025). Textual generation steps are clustered in semantic embedding space, forming cognitive states linked by directed edges. Dynamic trust scores and entropy metrics flag deadlock states ("cognitive vortexes"), and guided interventions (prompt hints, temperature perturbations) efficiently break reasoning loops, increasing accuracy and drastically reducing inference time, notably in small LLMs (Yan, 20 Dec 2025).

5. Integration with Episodic Memory and Vision–LLMs

In the VLEM (Vision–Language Episodic Memory) paradigm, the hippocampus is externalized as a trio of attractor subnetworks (CA3-like) governing "where," "what," and "when" dimensions, consolidating multimodal embeddings produced by a fixed vision–language neocortical encoder (e.g., CLIP) (Li et al., 7 May 2025). Prefrontal working memory modules (RNN slots) and a cross-attentional entorhinal gateway coordinate context transfer.

The attractor network is trained end-to-end with losses enforcing state-matching, contrastive separation, and input/output consistency, yielding robust, noise-resistant embedding retrieval and interpretable mapping of events and agent trajectories. Event recall from partial cues, graceful degradation under noise, and explicit recovery of spatial semantics demonstrate applicability to continual-learning agents and embodied simulation environments (Li et al., 7 May 2025).

6. Structural and Morphometric Models

HippMetric exemplifies the application of the external hippocampus concept at the structural and morphometric level, introducing a skeletal (s-rep) coordinate system aligned with hippocampal anatomy and lamellar architecture (Gao et al., 22 Dec 2025). The pipeline constructs a medial sheet parameterized along longitudinal (head-to-tail) and lamellar (cortical-thickness) axes, with boundary points recovered via radii ("spokes") from the skeleton.

This representation enables precise pointwise correspondence across subjects and over time, robust to the brain's complex folding variability. Quantitative evaluation demonstrates improved geometric accuracy, superior test-retest reliability (ICC for width/length >0.91, subfield ICC >0.7), and substantial gains in clinical discrimination and progression prediction (AUC up to 0.861 for Alzheimer's conversion). The methodology extends directly to other genus-zero neuroanatomical structures (Gao et al., 22 Dec 2025).

7. Memory Consolidation, Dynamics, and Biological Plausibility

A coupled neural field model elaborates the dynamics of hippocampal-to-neocortical consolidation, capturing why hippocampal memories are encoded and recalled rapidly, yet gradually rendered obsolete as neocortical engrams become self-sufficient (Moyse et al., 3 Apr 2024). Separate neural fields with distance-dependent synaptic plasticity, adult neurogenesis, spike-frequency adaptation, and synaptic depression account for the cortical–hippocampal two-speed regime. Hippocampal replay drives rapid memory bump formation, while neurogenesis ultimately erases hippocampal traces, leaving stable cortical representations (Moyse et al., 3 Apr 2024).

This two-speed learning framework—repeated across algorithmic, dynamical, and hardware levels—substantiates the external hippocampus as a computational substrate that is distinct, but deeply interfaced with, slow, robust perceptual and semantic memory systems.

---

In summary, the External Hippocampus Framework unifies diverse models in computational neuroscience, AI, robotics, and morphometrics. It operationalizes the hippocampus as a high-capacity, rapidly updated, interference-prone associative memory external to slower neocortical modules, implemented through a spectrum of data-structural, dynamical, topological, and anatomical paradigms, each with domain-specific adaptions and guarantees. These approaches collectively explain and reproduce key empirical phenomena in invariant perception, episodic recall, sequence memory, spatial navigation, cognitive planning, and disease biomarker quantification.