[Paper Review] The molecular memory code and synaptic plasticity: a synthesis

Samuel J. Gershman|arXiv (Cornell University)|Sep 11, 2022

Neural dynamics and brain function4 citations

TL;DR

This paper proposes a dual-memory framework in which intracellular molecules store generative model parameters (facts), while synapses store inference model parameters for fast, approximate Bayesian inference. By integrating molecular memory with synaptic plasticity under a free energy minimization principle, the theory reconciles long-standing conflicts between synaptic and molecular theories of memory.

ABSTRACT

The most widely accepted view of memory in the brain holds that synapses are the storage sites of memory, and that memories are formed through associative modification of synapses. This view has been challenged on conceptual and empirical grounds. As an alternative, it has been proposed that molecules within the cell body are the storage sites of memory, and that memories are formed through biochemical operations on these molecules. This paper proposes a synthesis of these two views, grounded in a computational theory of memory. Synapses are conceived as storage sites for the parameters of an approximate posterior probability distribution over latent causes. Intracellular molecules are conceived as storage sites for the parameters of a generative model. The theory stipulates how these two components work together as part of an integrated algorithm for learning and inference.

Motivation & Objective

• To resolve the longstanding conflict between synaptic plasticity and molecular memory theories in neuroscience.
• To explain the functional role of synaptic plasticity, which is often assumed to be essential but whose precise computational purpose remains unclear.
• To propose a biologically plausible mechanism where intracellular molecular processes encode long-term facts (e.g., quantities of space, time, number), while synapses encode parameters for efficient inference.
• To unify two distinct forms of memory—representational (facts) and computational (inference parameters)—under a single optimization principle: free energy minimization.
• To challenge the entrenched assumption that synapses are the sole site of long-term memory storage, advocating instead for a dual-system model grounded in predictive coding and Bayesian inference.

Proposed method

• Formalize memory as a two-part system: intracellular molecules store parameters of a generative model (e.g., beliefs about environmental variables), while synapses store parameters of an inference model (i.e., approximate posterior distributions).
• Frame learning and inference as a free energy minimization problem, where the brain optimizes both generative and inference model parameters to reduce prediction error.
• Model synaptic plasticity as Hebbian learning rules that update inference model parameters to improve predictive accuracy and reduce free energy.
• Propose that molecular plasticity—via epigenetic mechanisms like DNA methylation and histone modification—updates the parameters of the generative model, encoding long-term facts.
• Integrate the framework with predictive coding: feedforward pathways convey prediction errors, feedback pathways convey predictions, and both are implemented via layered cortical circuits.
• Use a computational theory of mind to show how the synergy between molecular and synaptic mechanisms enables efficient, adaptive behavior through approximate Bayesian inference.

Experimental results

Research questions

• RQ1 What is the true biological substrate of long-term memory storage for factual knowledge (e.g., quantities, durations, spatial relations) if not synapses?
• RQ2 Why does synaptic plasticity persist if it does not store memory content, given its high metabolic cost and strong behavioral correlates?
• RQ3 How can intracellular molecular mechanisms like epigenetic modifications serve as a stable, long-term memory code?
• RQ4 In what way do synaptic and intracellular processes jointly optimize the brain’s ability to perform inference and prediction?
• RQ5 Can a unified theory of memory be constructed that reconciles the generative model (molecular) and inference model (synaptic) under a single objective function?

Key findings

• Synapses are unlikely to be the primary site of long-term storage for representational content (e.g., facts about space, time, number), challenging the textbook synaptic plasticity model.
• Intracellular molecular mechanisms—particularly epigenetic modifications such as DNA methylation and histone acetylation—are plausible candidates for storing the parameters of a generative model of the world.
• Synaptic plasticity serves not memory storage per se, but the optimization of inference model parameters that enable fast, approximate Bayesian inference.
• The dual system—molecular for generative models, synaptic for inference models—works synergistically to minimize free energy, a core objective in predictive coding and active inference.
• The theory explains the transition from procedural learning (e.g., counting) to memory retrieval (e.g., instant recall) as a shift from computation to stored results, consistent with cognitive development data.
• The framework accounts for why synaptic plasticity is metabolically costly yet functionally essential: it enables rapid belief updating and prediction, not fact storage.