Memory consolidation and representational drift

This paper presents a phenomenological dynamical systems model that reinterprets memory consolidation as a brain-wide process of distributed engram trajectories, offering a functional explanation for representational drift as a mechanism that enhances long-term memory retention despite apparent instability in neuronal subsets.

The Gist

The Big Picture: How We Keep Memories Without Losing Our Minds

Imagine your brain is a massive, bustling library. Every time you learn something new, a librarian (your brain) writes a note about it and puts it on a temporary shelf near the entrance (the Hippocampus). This is where new memories live initially.

But here's the problem: The entrance shelf is small and crowded. If you keep putting new notes there, the old ones get pushed off and forgotten. To keep a memory forever, the librarian has to move it to the main archives deep inside the building (the Cortex). This process of moving memories from the "temporary shelf" to the "permanent archive" is called Memory Consolidation.

This paper asks a tricky question: How does the memory change while it's being moved?

The Main Idea: The "Drifting" Memory

For a long time, scientists thought memories were like static photos. Once you took a picture of a face, that picture stayed exactly the same forever.

However, recent experiments showed something weird. If you look at the specific neurons (brain cells) that hold a memory today, and then look at them again in a few weeks, the exact same cells aren't doing the work anymore. Different cells have taken over. The memory is still there, but the "staff" holding it has completely changed. This phenomenon is called Representational Drift.

Scientists used to think this drift was just random noise—like a glitch in the system or a mistake. They thought, "Oh, the brain is just getting sloppy."

This paper argues something different: The drift isn't a mistake. It's a feature. It's a deliberate, organized dance to keep memories safe.

The Analogy: The "Musical Chairs" of Memory

Imagine a game of Musical Chairs, but instead of people sitting on chairs, the "memory" is a song being passed around a circle of musicians.

1. The Classical View (Hippocampus to Cortex):
   Imagine the song starts with a violinist (Hippocampus). Over time, the song is passed to a cellist, then a drummer, and finally to a whole orchestra (Cortex). Once the orchestra has the song, the violinist stops playing it. The song has moved from one place to another.

2. The New View (Distributed Drift):
   The authors suggest it's not just a hand-off from one group to another. Instead, the song is being played by a huge, global orchestra all at once.

   • The melody (the memory) stays the same.
   • But the specific instruments playing the notes are constantly shifting.
   • One day, the violins play the high notes; the next week, the flutes take over.
   • Why? Because if the song is spread across thousands of instruments, it's much harder to lose the whole song if a few instruments break or get tired. By constantly shuffling who plays what, the memory becomes robust and long-lasting.

The "Hidden" Chaos: Why It Looks Random

Here is the clever part of the paper.

If you are a scientist looking at the brain, you can't see all the neurons at once. You can only see a tiny fraction of them (maybe 100 out of 100,000). This is called Subsampling.

• The Full Picture: The whole orchestra is playing a perfect, predictable, organized song. The shift in instruments is a planned, mathematical sequence.
• The Partial Picture: You are only watching 3 violinists. Because the other 99,997 musicians are shuffling the music behind the scenes, those 3 violinists look like they are playing randomly. They seem to be making mistakes or changing their tune for no reason.

The Paper's Conclusion: What looks like "random noise" or "drift" to us is actually a highly organized, deterministic process happening in the parts of the brain we can't see. The "randomness" is an illusion caused by our limited view.

Why Does This Matter?

1. It explains "Forgetting" vs. "Learning": The model shows how we can forget the boring details (like what you had for lunch on a Tuesday three years ago) but keep the important "gist" (that you love pizza). The brain actively shuffles the memory, keeping the important parts and letting the unimportant parts fade away.
2. It solves the "Stability vs. Plasticity" problem: How can the brain learn new things (plasticity) without erasing old things (stability)? By constantly moving the memory around to new neurons, the brain keeps the memory fresh and adaptable without losing it.
3. It changes how we study the brain: If we think the drift is just random noise, we might try to "fix" it. But if the drift is actually a smart strategy for long-term storage, we need to change how we record brain activity. We need to look at the whole picture, not just a tiny slice, to understand how memory really works.

Summary in One Sentence

Memory consolidation isn't just moving a file from one folder to another; it's a continuous, organized reshuffling of the entire brain's staff to ensure the story of "you" never gets lost, even if it looks like chaos when we only peek at a few cells.

Technical Summary

1. Problem Statement

The paper addresses two major gaps in understanding long-term memory:

• Systems Consolidation vs. Cellular Dynamics: While systems consolidation is classically described as the redistribution of memories from the hippocampus to the neocortex over time, the specific cellular mechanisms and how neuronal engrams (the physical traces of memory) evolve during this process remain poorly understood.
• Representational Drift: Experimental data shows that neuronal representations of stable tasks (e.g., place cells or sensory tuning) change continuously over days ("representational drift"), even when behavior remains stable. It is unclear whether this drift is a stochastic epiphenomenon (noise) or a directed, functional process.
• The Link: The authors propose that the active, directed process of memory consolidation might be the underlying cause of representational drift, and that the apparent randomness of drift in experiments is an artifact of observing only a small subset of neurons (subsampling).

2. Methodology

The authors developed a phenomenological dynamical systems model of memory engram dynamics.

• Core Model: The model treats memory engrams ($e$) as vectors representing the participation of neurons in a memory. The dynamics are governed by a linear differential equation:
  $$\frac{d}{dt}e = -\tau^{-1}e + Ae$$
  • $-\tau^{-1}e$: Represents passive forgetting (decay) with neuron-specific time constants.
  • $Ae$: Represents active consolidation, where the "consolidation matrix" $A$ redistributes the engram's activity among neurons over time.
• Readout: Behavioral output ($z$) is a linear readout of the engram: $z = A_{out}e$.
• Pattern Space Representation: To handle high-dimensional dynamics, the authors decompose the engram into a superposition of stable "patterns" ($p_\alpha$) with time-varying strengths ($w_\alpha$). This allows the system to be viewed as a sequence of pattern-to-pattern transformations rather than just area-to-area transfer.
• Simulation & Analysis:
  • The model was simulated using Euler integration.
  • Subsampling Analysis: The authors simulated a full population of neurons and then analyzed only a fraction (subsample) to mimic experimental limitations (e.g., calcium imaging of a few hundred neurons out of millions).
  • Mathematical Derivation: They used Green's function formalism and random matrix theory (specifically Gaussian random Hamiltonians) to analytically derive how subsampling transforms deterministic rotational dynamics into apparent stochastic decay.

3. Key Contributions

1. Unified Framework: The paper provides a mathematical bridge between the classical "hippocampus-to-cortex" view of consolidation and modern "distributed engram" views, framing consolidation as a trajectory through a space of brain-wide patterns.
2. Drift as a Feature, Not a Bug: It proposes that representational drift is not merely noise but a natural consequence of the deterministic, directed reorganization of engrams during consolidation.
3. The Subsampling Hypothesis: The authors demonstrate that deterministic dynamics can appear stochastic when observed through a subsampled lens. The "noise" observed in experiments is mathematically explained by the coupling between the observed subspace and the unobserved "bath" of neurons.
4. Power-Law Forgetting without Heterogeneity: The model achieves power-law forgetting (a hallmark of human memory) using a system with homogeneous neuronal forgetting times, simply by leveraging the sequential dynamics of distributed patterns with geometrically distributed time constants.

4. Key Results

A. Sequential Dynamics and Selective Consolidation

• Classical View: The model reproduces the classic hippocampus-to-cortex transfer as a feedforward network where an initial hippocampal pattern decays while a cortical pattern grows.
• Distributed View: By generalizing to distributed patterns, the model shows that memories can follow complex trajectories across the whole brain.
• Selectivity: Using a low-rank consolidation matrix, the model explains selective forgetting. Only components of an engram aligned with the "selection vectors" of matrix $A$ are consolidated; orthogonal components (e.g., episodic details) are passively forgotten, while aligned components (e.g., semantic gist) are retained. This explains semantization.

B. Stable Memories via Null-Space Dynamics

• The model demonstrates that a memory readout ($z$) can remain perfectly stable even as the underlying engram ($e$) changes drastically.
• This occurs if the consolidation dynamics occur entirely within the output-null space of the readout matrix ($A_{out}$). The engram rotates or drifts in dimensions that do not affect the behavioral output, preserving the memory while the neural representation changes.

C. Power-Law Forgetting

• By assigning geometrically distributed forgetting times to the sequential patterns ($\tilde{\tau}_\alpha$), the sum of these patterns decays as a power law.
• Crucially, the individual neurons in this system exhibit relatively homogeneous forgetting times (low coefficient of variation). This challenges previous models that required specific neuron- or area-specific learning rates to generate power-law forgetting.

D. Representational Drift and Subsampling

• Drift Hallmarks: The model reproduces key features of drift: gradual decorrelation of population vectors, shifting tuning curves, and the preservation of the "tiling" of stimulus space (the set of representations remains complete, just with different neurons).
• Deterministic vs. Stochastic: While the full system dynamics are deterministic, subsampling makes them appear stochastic.
  • When only a fraction of neurons is observed, the influence of the unobserved population acts as a "bath," causing the observed dynamics to decay and lose predictability.
  • Decoder Performance: A linear decoder trained on a subsampled population fails over time, whereas a decoder trained on the full population remains stable. This explains why drift is often interpreted as a failure of stability in experimental settings.
  • Fitting Failure: Attempts to fit linear dynamical systems to subsampled data fail to generalize to future time points (out-of-distribution), even if the underlying system is perfectly linear and deterministic.

5. Significance

• Reinterpretation of Drift: The paper shifts the paradigm of representational drift from a "problem to be solved" (noise to be compensated) to a "mechanism of memory maintenance." Drift is the visible signature of the brain actively reorganizing engrams to ensure long-term retention.
• Experimental Implications: It suggests that the apparent randomness in neural recordings may be an artifact of limited observability (subsampling). To truly understand consolidation dynamics, future experiments may need to record larger populations or use model systems (like Drosophila) where full population recording is feasible.
• Theoretical Unification: It unifies disparate phenomena—systems consolidation, selective forgetting, semantization, power-law forgetting, and representational drift—under a single dynamical systems framework based on linear recurrent networks and pattern-space trajectories.

In summary, the authors argue that memory consolidation is a directed, brain-wide redistribution of engrams. The "drift" observed in neurons is the necessary geometric consequence of this redistribution, which only appears stochastic due to the practical limitations of observing only a small fraction of the brain's neurons.