Engram: Cellular & Computational Memory

• Engram is a persistent memory trace comprised of sparse neuronal ensembles and compressed computational representations that enable recall.
• It bridges cellular mechanisms like synaptic plasticity with distributed circuit dynamics, highlighting its role in both biological and AI systems.
• Mechanistic insights such as Hebbian plasticity, synaptic scaling, and replay dynamics underpin the robust storage and retrieval of engrams.

Engram denotes the enduring trace by which an experience remains available for later reactivation. In contemporary neuroscience, it usually refers to a sparse population of neurons that is recruited during learning, modified by plasticity, and later reactivated during recall; in adjacent computational literatures, the same term has been extended to compressed latent representations, explicit associative memory structures, sparse task-specific subnetworks, and named memory phases in AI evaluation protocols (Szelogowski, 2 Jun 2025, Lucas, 2023, Liu et al., 4 Dec 2025). The concept is therefore both biologically specific and terminologically heterogeneous: it can designate a cellular memory substrate, a distributed circuit-level assembly, a repertoire of activation trajectories, or an engineered memory mechanism that borrows neurobiological language (Vladu et al., 7 Mar 2026).

1. Conceptual scope and historical usage

In the classical neuroscientific sense, the engram is the physical and biochemical trace of a memory in the brain. One influential formulation, traced to Richard Semon, treats memory as a “latent modification” left by experience and reawakened by cues during retrieval; modern work operationalizes this as a sparse population of neurons and synapses that encode, stabilize, and reconstruct remembered information (Szelogowski, 2 Jun 2025). This usage is narrower than generic “memory” and broader than a single synapse or single neuron: the same literature emphasizes an engram complex, meaning that memory is often distributed across multiple interconnected engram ensembles in different brain regions rather than confined to a single locus (Szelogowski, 2 Jun 2025).

The term is also used at different scales of description. Some authors describe an engram as a memory fragment or memory “chunk” that can be retrieved, chained, and replayed, so that “engrams become thoughts (awake) or dreams (in REM sleep)” (Spencer, 2024). Others explicitly reject the notion of a static stored snapshot and define an engram as a distributed recurrent network encoding a structured repertoire of possible activation trajectories, with dynamics summarized as $X_{t+1} \in T_i(X_t)$ for $E_i=(V_i,A_i,T_i)$ (Vladu et al., 7 Mar 2026). This wider systems-level usage preserves the core idea of a durable memory substrate while shifting emphasis from storage to executable structure.

No consensus exists on the exact physical implementation of an engram. One computational-neuroscience account frames the unresolved alternatives explicitly: engrams may be neuronal, synaptic, dendritic, or distributed across wider circuit-level organization (Lucas, 2023). This unresolved ontology is central rather than incidental, because many downstream debates—about sparsity, localization, indexing, consolidation, and recall—turn on what exactly is presumed to persist.

2. Cellular substrates, plasticity, and sparsity

Experimental engram research identifies engram cells through immediate early gene labeling such as c-fos and Arc, then probes necessity and sufficiency with optogenetics, chemogenetics, calcium imaging, and electrophysiology (Szelogowski, 2 Jun 2025). In this framework, neurons recruited during learning undergo lasting molecular and structural changes, and later retrieval depends on partial cue-driven reactivation of enough of the original sparse pattern to support pattern completion. The same work emphasizes that recruited engram cells are not static labels: in CA3 pyramidal neurons, increased excitatory input is followed, over about a day, by increased intrinsic excitability, while inhibitory inputs can also increase, implying a stabilizing excitatory-inhibitory balance rather than unbounded potentiation (Szelogowski, 2 Jun 2025).

The canonical mechanistic account remains Hebbian. A representative update rule is

$\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$

where $\omega_{ij}$ is synaptic weight, $u_j$ presynaptic activity, and $v_i$ postsynaptic activity (Szelogowski, 2 Jun 2025). On its own, however, pure Hebbian strengthening is unstable. A stabilized formulation combines Hebbian plasticity with synaptic scaling: $$\frac{d\omega_{ij}}{dt}=\mu u_j v_i+\gamma(v_T-v_i)\omega_{ij}^2.$$ This is used to argue that durable memory traces emerge from potentiation plus homeostatic regulation rather than from potentiation alone (Szelogowski, 2 Jun 2025).

Sparsity is treated as a defining computational virtue of engrams. Sparse ensembles reduce metabolic cost, increase effective capacity, lower overlap between memories, and improve interference resistance and pattern separation (Szelogowski, 2 Jun 2025). One reviewed result states that memory capacity can scale approximately as

$\left(\frac{N}{\log N}\right)^2$

for very sparse activity, though the same source stresses that optimal performance may occur under weak rather than maximal sparsity regularization (Szelogowski, 2 Jun 2025). In that sense, an engram is not merely “small”; it is selectively sparse in a way that trades representational richness against interference.

3. Distributed organization, replay, and cross-modal expansion

At the circuit and systems levels, engrams are increasingly described as distributed and dynamically reconfigurable. A Drosophila study reports that multisensory learning recruits visually selective mushroom body Kenyon cells into an olfactory memory engram, so that cross-modal binding expands the neurons supporting later recall (Okray et al., 30 Apr 2026). In that account, multisensory appetitive and aversive training improves memory for the combined cue and for each cue alone, DPM transmission is uniquely required during multisensory memory formation and later enhanced olfactory memory expression, and the broadened engram permits a single sensory feature to retrieve a multimodal experience (Okray et al., 30 Apr 2026). The claim is explicitly not that multisensory learning merely strengthens a fixed olfactory trace, but that it enlarges the neuronal population constituting the trace.

Replay-based perspectives extend the engram beyond overt recall. One evolutionary account describes an “unconscious engram playback and discrimination” mechanism in birds and nearly all mammals, recognized by its side-effect: dreams (Spencer, 2024). In that model, sleep reactivates some new and some old engrams, strings them into sequences, and discriminates sensible from nonsensical associations, yielding “more brain capability at a given size” through storage optimization (Spencer, 2024). The engram is therefore not only a retrieval substrate but also a unit of offline reorganization.

DIME generalizes this systems view by distinguishing engrams from execution threads and hyperengrams. Engrams are the representational substrate; execution threads are the temporally extended activation paths traversing that substrate; hyperengrams are large-scale, value-marked integrative states associated with operational conscious access (Vladu et al., 7 Mar 2026). This architecture formalizes a recurring theme in engram research: a memory trace is durable not because it is inert, but because it supports constrained families of reactivation, replay, prediction, and reconsolidation.

4. Computational abstractions and formal analogies

Several machine-learning-oriented works recast engrams as compressed and indexed internal representations. One influential proposal models an engram through an autoencoder-like system comprising an encoding system, a latent space, a decoding system, and a layer of concept nodes that “connect[s] all value points in the latent space that are related to the same concept” (Lucas, 2023). In minimal notation, input $x$ is encoded to $z=f_{\text{enc}}(x)$, reconstructed as $E_i=(V_i,A_i,T_i)$0, and linked to a concept index $E_i=(V_i,A_i,T_i)$1 that binds multiple latent codes $E_i=(V_i,A_i,T_i)$2 across modalities (Lucas, 2023). Here the engram is not the reconstruction alone but the organization of compressed representations plus the index that associates them.

A related line of work argues that latent-space dimension constrains what an engram can represent. The key condition is $E_i=(V_i,A_i,T_i)$3, where $E_i=(V_i,A_i,T_i)$4 is the intrinsic dimension of the data manifold and $E_i=(V_i,A_i,T_i)$5 the latent dimension of the encoding space (Lucas, 2024). This treats the engram as a biologically implemented autoencoder over recurrent neural networks, with latent geometry determining abstraction, separability, and concept formation (Lucas, 2024).

Homeostatic and graph-theoretic models pursue different analogies. One proposal uses an XOR motif built from excitatory and inhibitory neurons as a biological analogue of a loss function: the circuit provides 0 when incoming and reference signals are equal and 1 otherwise, thereby generating a mismatch signal until learning stabilizes the representation (Lucas et al., 2024). Another defines the memory engram as a connected subgraph $E_i=(V_i,A_i,T_i)$6 in an active directed graph, with local node autonomy and distributed storage via node-internal index tables rather than a global controller (Wei et al., 2023). These models differ sharply in implementation but converge on the same abstraction: the engram is a structured, persistent carrier of information, not merely a transient activation pattern.

5. Deep-learning architectures and AI systems

In modern deep learning, “engram” has become a productive design metaphor for explicit memory. In continual learning, engramBNN treats an engram as a sparse, probabilistic memory trace instantiated by a Bernoulli-sampled binary gate $E_i=(V_i,A_i,T_i)$7 that selects a context-specific subset of units in a metaplastic binarized neural network (Aguilar et al., 27 Mar 2025). The method is reported as the only strategy reaching average accuracies above 20% in the class-incremental setting, while also achieving peak GPU and RAM usage of 4.51% and 19.51%, respectively (Aguilar et al., 27 Mar 2025). In recurrent modeling, the Engram Neural Network introduces an explicit memory matrix $E_i=(V_i,A_i,T_i)$8 and Hebbian trace $E_i=(V_i,A_i,T_i)$9, with retrieval through content-based attention over $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$0; performance is broadly comparable to RNN, GRU, and LSTM baselines while making memory dynamics inspectable (Szelogowski, 29 Jul 2025). In neural cellular automata, EngramNCA separates public visible state from private cell-internal “gene” channels, using GeneCA and GenePropCA to express and propagate persistent private memory across cells (Guichard et al., 16 Apr 2025).

LLM work uses Engram for a different class of mechanisms: conditional memory via deterministic $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$1-gram lookup. One architecture modernizes classic $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$2-gram embedding with tokenizer compression, multi-head hashing, contextualized gating, and multi-branch integration, defining a sparsity-allocation trade-off between routed experts and memory tables (Cheng et al., 12 Jan 2026). That work reports a U-shaped scaling law in the allocation ratio $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$3, with the best results when about 20–25% of the sparse parameter budget is moved from experts into memory; at 27B scale, the memory-augmented model improves over an iso-parameter and iso-FLOPs MoE baseline on knowledge, reasoning, code, math, and long-context retrieval, including Multi-Query NIAH from 84.2 to 97.0 (Cheng et al., 12 Jan 2026). Later work argues that order-specific hash tables waste nested structure and proposes tensorized shared factors for $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$4-gram embeddings, reporting comparable or better validation performance than Engram-style modules with fewer parameters (Zhou et al., 6 Jun 2026). Another variant, Memory Grafting, replaces trainable memory values with frozen hidden states from a pretrained grafting model and uses exact longest-match suffix lookup with a hash-based Engram fallback; in a 2.8B-scale setting it improves average benchmark score from 51.95 for MoE and 52.43 for vanilla Engram to 53.86 (Cheng et al., 20 May 2026).

The same label is also used in settings where the underlying object is not a memory trace in the neuroscientific sense. In autoregressive image generation, a vision-adapted Engram module with 2D spatial $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$5-gram hashing does not improve FID over a pure AR baseline, and probing experiments indicate behavior closer to a gated architectural side-pathway than a strong content-addressed retriever (Wang et al., 13 May 2026). In multi-agent LLM privacy evaluation, Engram is the initial memory-seeding phase of MAMA: only the target agent receives $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$6, producing an initial state $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$7 before the Resonance phase propagates traces through the graph (Liu et al., 4 Dec 2025). In agentic system optimization, Engram names a sequence-of-agents architecture in which each run writes code snapshots, logs, and results to an Archive and distills high-level insights into a persistent Research Digest for the next fresh-context agent (Karimi et al., 22 Mar 2026). In long-term memory for conversational agents, the name labels external memory engines rather than neural substrates: one bi-temporal system reports 83.6% versus 73.2% for full-context answering on the full 500-question LongMemEval$\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$8, using about 9.6k versus 79k tokens, while another lightweight typed-memory system reports 71.40 versus 56.20 on LongMemEval with about 1.0–1.2K tokens per query instead of 101K (Wang, 5 Jun 2026, Patel et al., 17 Nov 2025).

6. Terminological drift, misconceptions, and open questions

A recurrent misconception is that an engram must be a single localized “memory cell.” The surveyed literature does not support that simplification. Biological accounts emphasize sparse populations, distributed engram complexes, and partial-cue pattern completion rather than one-cell storage (Szelogowski, 2 Jun 2025). Systems theories go further by treating engrams as recurrent structures with many admissible micro-trajectories, not fixed attractors (Vladu et al., 7 Mar 2026). Computational models likewise distribute memory across latent spaces, graph substructures, or explicit memory banks rather than a unique storage locus (Lucas, 2023, Wei et al., 2023).

A second source of confusion is terminological drift across AI. In some papers, “Engram” denotes a memory module; in others, a persistent archive, a typed retrieval layer, or merely an initialization phase (Cheng et al., 12 Jan 2026, Karimi et al., 22 Mar 2026, Liu et al., 4 Dec 2025). This suggests a family of related metaphors—durable trace, selective retrieval, persistent state, sparse activation—rather than a single cross-domain technical definition. The drift is methodologically useful but conceptually hazardous: identical terminology can refer to neuronal ensembles, hash-keyed $\frac{d\omega_{ij}}{dt}=\mu u_j v_i,$9-gram tables, bi-temporal fact stores, or privacy-seeding protocols.

The unresolved scientific questions remain substantial. One review lists as open problems how to determine exactly when an engram has been found, what information content an engram stores, how distributed engram complexes are organized across regions, how stable memories remain reconstructive and mutable, and how neuromodulation, glia, and diverse adaptation mechanisms should be integrated into models (Szelogowski, 2 Jun 2025). A related computational-neuroscience proposal states bluntly that “we have no answer” for the physical implementation of encoding in the brain, even if autoencoder analogies are useful (Lucas, 2023). In applied AI, mechanistic ambiguity persists as well: for example, image-generation results indicate that an Engram module can be functionally important while still failing to behave as true content-addressed retrieval (Wang et al., 13 May 2026).

Across these literatures, the durable core of the concept is narrower than its modern vocabulary. An engram is, in the strict sense, a persistent substrate that makes later memory reactivation possible. What varies is the assumed substrate—cells, synapses, dendrites, circuits, latent codes, subnetworks, lookup tables, or external stores—and the operational emphasis placed on it: storage, indexing, replay, consolidation, retrieval, or architectural persistence.