REM-like Sleep Reweights Memory Traces Instead of Treating Them Equally
A new preprint argues that REM sleep may do more than passively replay memories. In recurrent neural networks trained with Equilibrium Propagation, sleep-like reactivation reorganized competing memory traces so they interfered less with one another, creating a cleaner separation between similar representations.
The study asks a long-running question in neuroscience: why does REM sleep appear to help memory consolidation when the brain already spent wakefulness encoding the information? The answer proposed here is mechanical rather than rhetorical. During sleep, reactivation was not spread evenly across all stored patterns. It was dominated by the population encoding the most recent memory, and that asymmetry changed the synaptic balance in a way that reduced crosstalk.
In the model, the most recently recruited neuronal population became the main target of reactivation. Net excitatory input to that population fell, with the strongest reductions coming from populations tied to older memories. As those older inputs weakened, the neurons that represented prior memories could respond to their own cues again. The result was not simple forgetting or uniform synaptic decay. It was a selective reorganization that let overlapping representations occupy more distinct neural territory.
Targeted Pruning, Not Global Synaptic Reset
The authors describe this process as reactivation-targeted pruning. The term captures the core logic of the model: plasticity is concentrated in one active population, and synapses that stay silent while those neurons fire are the ones most likely to be reduced.
One sentence captures the central consequence. Equalizing the same total synaptic change across all target populations erased the recovery effect.
That finding matters because it separates a specific sleep mechanism from a generic weakening effect. If sleep merely diluted synaptic strength across the board, the system would not recover the ability of older memories to drive their own neurons. Instead, the selective placement of plasticity appears to be the decisive factor. The network needed the imbalance created by reactivation, not just a net decrease in excitatory drive.
The model also linked this behavior to reduced interference between tasks. By decorrelating task-specific neuronal populations, REM-like sleep moved overlapping memories toward more orthogonal representations. In practical neural terms, that means the same network could preserve multiple similar experiences with less mutual contamination. The study frames this as a plausible route by which sleep converts unstable traces into more distinct stored patterns.
Excitation, Inhibition, and the Stability of Recalled Patterns
The work focused on the balance between excitation and inhibition, but the most striking result did not require elaborate inhibitory remodeling. The mechanism held under Dale’s law, preserving the principle that a neuron’s synapses are either excitatory or inhibitory. Even more notable, depression of excitatory synapses alone was sufficient to reproduce the effect, without any inhibitory plasticity.
That is a meaningful constraint. It suggests that the memory-separating effect of REM-like sleep can emerge from changes in excitatory connectivity alone, provided those changes are shaped by reactivation rather than distributed indiscriminately. The network does not need a broad rewiring of inhibition to accomplish the separation. It needs selective weakening of the inputs that are irrelevant at the moment a population is active.
The mechanism also fits a simple biological intuition. When a memory trace is reactivated, the synapses that fail to participate are positioned to be downscaled. Over repeated cycles, this could make the active trace more self-contained while reducing the overlap with older traces that once competed for the same neurons. The model therefore links a computational principle—orthogonalization—to a plausible synaptic process that can operate during sleep.
What the Network Model Can and Cannot Claim
This is a computational study, not a human sleep experiment. Its value lies in showing that a biologically plausible learning system can reproduce a specific kind of memory refinement without relying on unrealistic assumptions. Equilibrium Propagation gives the model a closer tie to known neural dynamics than many abstract machine-learning rules, and the authors use that framework to argue that REM-like sleep can support consolidation through targeted synaptic rebalancing.
The preprint also points toward a broader design principle for artificial networks. If memory interference can be reduced by restricting plasticity to the population currently reactivated, then sleep-inspired algorithms may improve continual learning systems that otherwise overwrite older information. That implication sits outside bedside care, but it has clear relevance for how brain-inspired networks are built and trained.
Still, the study remains at the level of mechanism inference. It does not establish how faithfully the model matches the full complexity of REM sleep in the intact brain, nor does it show direct experimental confirmation in animals or people. The work instead narrows the discussion to a specific proposal: reactivation during REM-like sleep may preserve memory by preferentially weakening silent inputs onto the most recently active population, allowing older representations to regain their own functional space.
In this preprint, the central result is not that sleep strengthens everything equally. It is that sleep may sharpen separation by acting unevenly, and by doing so, it can turn overlapping traces into more independent ones.