The Mechanics of Dynamic Representational Drift

20 concepts8 papers1 proposal

Traditional neuroscience frameworks long treated memory engrams as static snapshots—fixed ensembles of neurons locked together to preserve a specific experience over time. However, empirical observations reveal a stark contradiction: only a small fraction of the neurons active during initial learning remain active during subsequent recall. This phenomenon, known as representational drift, demonstrates that physical memory traces are highly volatile, undergoing continuous turnover during consolidation and storage.

How can the brain maintain stable, retrievable memories when the underlying neural substrate is constantly shifting? A growing body of computational and experimental research suggests that this drift is not random noise. Instead, it is a structured, self-regulating mechanism that optimizes memory organization, flexibility, and temporal tracking.

The Cellular Engine: Intrinsic Excitability Fluctuations

To understand why representations drift without destroying the memory itself, we must look at the biophysical properties of individual neurons. Neural assemblies are not hardwired, permanent circuits; their membership is fluid and stochastic.

As modeled by Geoffroy Delamare et al. in Drift of neural ensembles driven by slow fluctuations of intrinsic excitability, individual neurons experience slow, spontaneous fluctuations in their intrinsic excitability over time. When a memory is reactivated or a stimulus is re-encountered, the network preferentially recruits neurons that currently happen to have a lower threshold for firing. This creates a competitive, rolling membership for the memory ensemble.

Time 1: [A] [B]  C   D   (Neurons A and B are highly excitable -> Engram)
Time 2:  A  [B] [C]  D   (Excitability shifts; B and C are recruited -> Drifted Engram)
Time 3:  A   B  [C] [D]  (Excitability shifts further; C and D form the Engram)

This gradual turnover does not cause the network to collapse into runaway activity or lose its structure. In Purely STDP-based assembly dynamics: Stability, learning, overlaps, drift and aging, Paul Manz et al. demonstrate that these intrinsic firing rate fluctuations allow individual neurons to transition smoothly between different functional assemblies. Crucially, spike-timing-dependent plasticity (STDP) allows the overall structural connectivity of the network to remain stable even as individual cellular components turn over.

As a result, the physical trace of a memory evolves significantly during the hours and days following initial learning rather than remaining fixed, a process further detailed by Douglas Feitosa Tomé et al. in Dynamic and selective engrams emerge with memory consolidation.

Turning Drift into a Clock: Temporal Tagging

If memory representations are constantly changing, why hasn't evolution eliminated this drift? The answer lies in a powerful computational byproduct: drift naturally encodes time.

Because excitability fluctuations occur gradually, the composition of a neural ensemble shifts incrementally. If two events happen close in time, they will recruit highly overlapping sets of neurons because the brain's baseline excitability landscape hasn't changed much. If they happen days apart, the landscape will have shifted completely, leading to distinct ensembles.

Geoffroy Delamare et al. showed that this gradual drift provides an inherent "temporal tag." Downstream networks can decode the chronological order of past events simply by measuring the degree of overlap between ensembles, eliminating the need for an explicit digital timestamp.

An alternative, macro-level mechanism for this temporal tracking involves physical migration across neural layers. In Computational models of episodic-like memory in food-caching birds, Nicola S. Clayton demonstrates that temporal information can be mapped onto spatial properties via systems consolidation. As memories age, they migrate through a sequence of neural layers in a pipeline. The system can decode when an event occurred simply by monitoring which layer is active during retrieval, transforming memory age into a physical coordinate.

Generalization and Adaptive Decision-Making

Beyond tracking time, dynamic representations prevent memory systems from becoming overly rigid. If an engram were a static snapshot, it would overfit to the exact conditions of the initial experience, rendering it useless for predicting outcomes in slightly different contexts.

Early foundational work in Organization and Retrieval in a Conceptual Memory noted that memory systems must continuously reorganize and generalize information as new events are added to maintain organizational coherence and prevent an index explosion. This insight is formalized in the predictive frameworks of Adaptive Stimulus Representations: A Computational Theory of Hippocampal-Region Function, which shows that dynamically constructing and adapting stimulus representations enhances discrimination learning and generalization in complex or evolving environments.

This continuous modification is highly active in the hippocampal formation. As reviewed by Aidan J. Horner et al. in Plasticity of hippocampal memories in humans, hippocampal representations undergo constant trial-by-trial strengthening and weakening. This plasticity ensures that memories remain flexible and relevant for ongoing decision-making.

Indeed, Erie D. Boorman et al. in Two Anatomically and Computationally Distinct Learning Signals Predict Changes to Stimulus-Outcome Associations in Hippocampus demonstrate that these rapid updates to stimulus-outcome associations are driven by distinct anatomical learning signals, allowing the brain to dynamically adjust its internal model of the world to guide future actions. Representational drift, therefore, is not a flaw of a leaky biological system—it is the mechanism that makes memory functional, adaptive, and deeply integrated with time.

Key Foundational Work

Organization and Retrieval in a Conceptual Memory for Events or CON 54, Where Are You? (1981) — contributed 2 concept elements to this direction.
Adaptive Stimulus Representations: A Computational Theory of Hippocampal-Region Function (1992) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_rate_driven_representational_drift_3de413fd,
  title = {rate-driven representational drift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rate-driven-representational-drift/rdr_3de413fdcd704fe6b1a804e829a8c8cd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_rate_driven_representational_drift_3de413fd_v1,
  title = {rate-driven representational drift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rate-driven-representational-drift/rdr_3de413fdcd704fe6b1a804e829a8c8cd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake