Replay Makes Memories Rigid

01The evidence

• Hippocampal replay and network dynamics have been analysed with topological methods that expose a small number of dominant states, and the transitions between them resemble a system moving between metastable configurations rather than a system storing items in slots.Topological analyses of hippocampus

02The tension

Replay quantity correlates with consolidation only loosely, and in some conditions more replay produces more interference. If replay were simply transferring or strengthening traces, the correlation should be tight. The field has a mechanism with a weak dose-response.

03The theory

Think of a memory as a set of satisfied constraints among neurons rather than as a pattern of weights. Learning adds constraints; replay revisits them, and each revisit increases the number of constraints satisfied simultaneously. When enough are satisfied, the network becomes rigid: perturbing a few synapses no longer changes the stored pattern.

This is a glass transition, and glasses have a measurable order parameter. Rigidity can be probed by applying a small perturbation and measuring how far the state moves. A freshly encoded memory is soft and easily nudged. A consolidated memory is stiff, and its stiffness is a direct measure of the energy barrier protecting it.

The reframing explains several awkward results. More replay can hurt if it drives the network into a different, competing basin. Interference is a melting event, not overwriting. And retention should be predicted by the barrier height rather than by the strength of the weights, which is a testable difference: two memories with identical weight magnitudes but different constraint satisfaction should have different retention.

The prediction that matters is temporal. Glass transitions show critical slowing down as they approach. If consolidation is one, the relaxation time of the network should lengthen in the hours before a memory becomes stable, which is measurable in the return to baseline after perturbation.

04What this is not

05Predictions

• P1Measured stiffness, the response to a calibrated perturbation, should rise overnight and predict twenty-four-hour retention better than synaptic weight change.
• P2Random stimulation after learning, a reheating operation, should increase forgetting of the consolidated memory while leaving the just-encoded one intact.
• P3The network should show critical slowing down, a lengthening relaxation time, before the transition.

06What would refute it

If retention is fully predicted by weight magnitude and stiffness adds nothing, the landscape account is decorative rather than explanatory.

07What it would change

It gives memory research an order parameter and a reason to measure stability rather than strength, which changes how consolidation experiments are designed.