What should we forget? A computational model of memory consolidation

Technical Summary: A Computational Model of Memory Consolidation via Predictive Coarse-Graining

Problem Statement
Biological memory systems, specifically neural and adaptive immune systems, face a fundamental computational constraint: future situations rarely repeat past experiences exactly. Consequently, memory cannot function by storing exact copies of episodes. Instead, it must solve a compression problem where the target is not the reconstruction of the original experience, but the preservation of distinctions that alter future responses. The system must discard irrelevant variation while retaining "consequential differences" to support generalization, all while operating under physical constraints regarding storage capacity and maintenance costs.

The paper formalizes this challenge as scaffold-flow memory. In this architecture:

• Flow: Represents fast, state-dependent responses (e.g., immediate neural firing or immune cell activation).
• Scaffold: Represents slowly changing physical variables that constrain future dynamics (e.g., synaptic weights, receptor affinities, clone abundances).
• Consolidation: The process of writing a "predictive coarse-graining" of experience into the scaffold.

The core problem is to identify a grouping of experiences that satisfies four simultaneous requirements:

1. Behavioural Sufficiency: Experiences may be merged only if they induce the same distribution over futures.
2. Generalization: The grouping must assign novel variants to existing categories if they share the same future, while excluding variants requiring different responses.
3. Non-leaky Abstraction: Once an experience is labeled, its coarse dynamics must be predictable from the label alone, without dependence on discarded microscopic details.
4. Affordability: The grouping must be writable and maintainable in physical structure at an acceptable cost.

Methodology
The author proposes a computational framework where consolidation acts as constrained model selection over both coarse-grainings and their physical realizations. The methodology involves:

1. Theoretical Formalization:

   • Defining Leakage ($\Lambda_T$) as a measure of failure in coarse autonomy. If a grouping $\pi$ assigns states $x$ and $x'$ to the same group, but their futures diverge, the abstraction is "leaky." The paper derives a bound showing that excess future error is proportional to leakage.
   • Establishing a Recoverability Trichotomy (Proposition 1) based on the mixing time ($\tau_{mix}$) of the experience process relative to the agent's lifetime ($T$):
     • Case 1 (Ergodic, fast mixing): Structure is recoverable; storage buys speed (memoization).
     • Case 2 (Ergodic, slow mixing): Structure is not recoverable within a lifetime; storage is mandatory (inscription).
     • Case 3 (Non-ergodic/Drifting): No convergent target exists; storage is the only mechanism of retention.
   • This distinction argues that consolidation is not optional housekeeping but a necessity for systems facing non-self-averaging or drifting environments.

2. Simulation Models:
   The framework is tested across three distinct substrates to demonstrate substrate-independent principles:

   • Recurrent Neural Network: A network of $N=500$ rate units storing sparse patterns. Consolidation is modeled as increasing recurrent gain ($g$), deepening basins of attraction. Spectral analysis (transfer operator theory) is used to identify metastable categories as near-unit eigenvalues.
   • Associative Memory: A Willshaw model and a Modern Hopfield network. The study compares episodic storage against predictive grouping (quotients). It specifically investigates the separation of keys (cues) and values (responses), allowing for coarse-graining on the value side while preserving fine cue addresses.
   • Stochastic Immune Model: A germinal-center affinity maturation simulation. Grouping is not pre-defined but emerges through mutation and selection. A selection exponent ($p$) controls the trade-off between clones with strong single-strain matches versus broad coverage. The model includes a finite repertoire budget to penalize over-refinement.

3. Ablation Studies:
   To test the necessity of the scaffold, the author simulates "lesions" (deletion of scaffold entries) under different generative regimes (stationary short-recurrence, singular, and drifting). This tests whether experience can regenerate lost structure based on the recoverability trichotomy.

Key Results

• Emergence of Slow Modes: In the recurrent network, consolidation (increasing gain) converts fast fluctuations into persistent categories. Spectral analysis confirms that consolidated categories appear as slow modes (eigenvalues near unity) with distinct gaps from the rest of the spectrum. The lifetime of these categories follows Arrhenius-Kramers scaling, and first-passage times predicted by the transfer operator match measured values.
• Optimal Granularity is Intermediate: Across all three systems, future risk (error) exhibits a U-shaped relationship with grouping granularity.
  • Over-fine granularity: Wastes capacity, fails to generalize to novel variants, and leads to poor estimation of future statistics due to sparse data.
  • Over-coarse granularity: Merges experiences requiring different responses, leading to false alarms (neural) or autoimmunity (immune).
  • Optimum: The minimum risk occurs at an intermediate granularity, typically close to the number of truly distinct futures ($r^*$). In the immune model, the optimum is slightly higher than $r^*$ to account for within-family variability.
• Asymmetric Costs of Error: The two arms of the U-curve represent distinct failure modes. Over-merging violates behavioral sufficiency (causing self-reactivity or false recall), while over-splitting violates generalization (causing missed opportunities or lack of coverage). The optimal balance shifts based on environmental variability and the relative cost of false alarms versus misses.
• Recoverability and Lesion Recovery: Ablation studies confirm the theoretical trichotomy.
  • In stationary, short-recurrence regimes, deleted scaffold entries are regenerated by subsequent experience without retraining.
  • In singular regimes (events that never recur), deleted entries are permanently lost.
  • In drifting regimes, the system "relearns" the new world, but the original encoded content is lost. This demonstrates that relearning is not recovery; the loss is real but not lesion-attributable because the target no longer exists.
• Value-Side Coarsening: In the Modern Hopfield model, the paper demonstrates that predictive grouping is most effective when applied to the values (responses) rather than the keys (cues). This allows the system to retain fine-grained cue discrimination while pooling responses for equivalent futures, significantly reducing storage requirements (e.g., storing 24 patterns vs. 1536) while minimizing risk.

Significance and Claims
The paper claims to identify a common computational principle underlying neural and immune memory consolidation, despite their disparate biological mechanisms. The central thesis is that memory consolidation is the selection of a predictive, dynamically usable, and affordable coarse-graining of the past.

Key contributions include:

1. Unification: Framing neural and immune memory as solving the same "scaffold-flow" problem, where fast responses are shaped by slow physical structures.
2. Necessity of Storage: Demonstrating that storage is not merely an efficiency device (memoization) but a mandatory mechanism (inscription) for systems operating in environments that are non-ergodic, slowly mixing, or drifting.
3. Predictive Granularity: Establishing that the optimal level of memory detail is not "as fine as possible" nor "as coarse as possible," but an intermediate point that balances behavioral sufficiency with generalization and cost.
4. Testable Predictions: The framework generates specific, falsifiable predictions, such as the U-shaped risk curve across substrates, the spectral signature of consolidated memories (slow modes), and the differential recovery of memories based on the recurrence time of the underlying statistics (recoverable vs. non-recoverable).

The author concludes that while neural and immune systems share no common molecules or equations, they share a computational solution: a coarse-graining that preserves consequential differences, supports autonomous coarse dynamics, and justifies its physical cost.