SYSTEMS MEMORY CONSOLIDATION AS AN ANTIDOTE TO REPRESENTATIONAL DRIFT

Neuronal representations of acquired memories are inherently dynamic, a phenomenon known as representational drift. Over days to weeks, hippocampal and cortical population codes drift as neurons enter and leave memory ensembles yet, recall and behaviour remain stable. This raises a fundamental question: how can stable behaviour arise from unstable neural representations? We asked whether systems memory consolidation can counter hippocampal representational drift and enable stable behaviour.
Here, we developed a two-region recurrent neural network to model the engram dynamics in the hippocampus (HPC) and anterior cingulate cortex (ACC). We also included a single output neuron as a fixed decoder. Each network comprised excitatory neurons with recurrent Hebbian plasticity and biologically inspired connectivity between HPC and ACC.
Simulations produced the gradual emergence of engram cells with strengthened recurrent connectivity. The HPC network quickly developed an engram ensemble that exhibited representational drift driven by excitability fluctuations. The model reproduced several experimental features of systems memory consolidation: for example, feedforward drive from HPC enabled the progressive formation and stabilisation of ACC engram cells. The model also recapitulated recent experimental findings on the time window for synaptic plasticity in cortex necessary for systems memory consolidation. Crucially, the fixed output neuron in our model maintained reliable decoding across time because the consolidated cortical representation compensated for hippocampal drift.
Our model demonstrates that systems consolidation can reconcile dynamic neural representations with stable behaviour. By transferring memories to cortex, consolidation provides a robust mechanism for maintaining consistent behavioural output despite ongoing hippocampal reorganisation.