Could sleep-inspired maintenance help AI agents manage their memories?

An AI agent that remembers earlier interactions faces a problem beyond finding somewhere to store them: deciding which memories deserve to survive. In Do agents need to sleep?, Ido Haber and Giulio Tononi propose borrowing a principle from biological sleep. Agents, they argue, may need periods of reduced environmental interaction to review accumulated experience and selectively reduce its burden. The preprint was first submitted on November 1, 2026, and revised on November 3. It presents a conceptual argument and a testable prediction, not a new experimental demonstration.

The authors focus on agents that retain and update memory across sessions. Drawing on earlier research, they describe several ways this can go wrong. Retrieved mistakes can influence subsequent answers; repeated summarization can move stored accounts away from their original meaning; irrelevant records can crowd retrieval; and contradictions can accumulate. In this account, performance can deteriorate even when the underlying language model has not changed, because the information supplied from memory has become less useful.

This is an instance of the stability–plasticity dilemma: a learning system must accommodate new experience without destabilizing what it already knows. Haber and Tononi argue that increasing storage or improving retrieval does not, by itself, resolve the underlying costs of continually accumulating and modifying memories.

A three-part model borrowed from sleep

Their starting point is the Synaptic Homeostasis Hypothesis, a framework in which waking learning produces a net strengthening of connections between neurons. That strengthening carries costs, including energy use, reduced signal-to-noise ratios and limits on further learning. Sleep, in this framework, allows selective weakening while protecting important circuitry.

The proposed bridge to AI comes from a selectionist account called “ringing, resonance and renormalization,” or RRR. Here, selectionist means that preservation emerges from competition among circuits rather than from explicit instructions deciding the fate of each connection. During ringing, spontaneous activity samples circuits. Resonance reflects how successfully that activity sustains itself through coordinated recurrent connections. Renormalization weakens connections while protecting those participating in coherent activity.

The authors cite biological evidence for sleep-associated reductions in synaptic strength or structure across species. That evidence motivates their proposal, but it does not establish that the proposed digital equivalent will improve AI agents.

For digital memory, ringing could mean randomly traversing stored traces and probing unfamiliar combinations independently of an active user query. Resonance could involve assessing semantic alignment among those traces, potentially using cosine similarity in a vector database. Renormalization could merge redundant records, prune low-importance material and compress detailed episodes into broader abstractions. These are suggested implementations, not a specified and validated algorithm.

The missing element, the authors say, is a unified cycle: broad sampling would feed an assessment of coherence, which would then guide selective compression. They identify renormalization as the phase most developed in existing memory-management practice.

Why withdraw from interaction?

Disconnection matters to their argument because incoming experience could change the memory store before its competing traces have been evaluated. Adding records or adjusting retrieval weights during consolidation would, they contend, disrupt that evaluation.

This need not mean switching an entire agent off for every maintenance operation. The paper suggests running RRR on an isolated snapshot of one memory partition while the agent continues answering queries. It also discusses LightMem, which separates lightweight online updates from a deferred offline reconciliation pass. Nevertheless, the authors argue that comprehensive consolidation would require a complete pause in environmental interaction. That requirement is part of their proposal, rather than an experimentally established result for AI systems.

The crucial comparison is with what they call instructionist management. Their example, Mem0, uses a language model to compare each new fact with similar stored memories and choose whether to add, update, delete or leave information unchanged. Such decisions are explicit and local, rather than emerging from competition across the store.

Haber and Tononi predict that selectionist consolidation will outperform instructionist alternatives as memories become larger and more interconnected. Their reasoning is that explicitly deciding what to preserve will become combinatorially difficult. The paper reports no benchmark comparison establishing that advantage. Its contribution is a research hypothesis: sleep-like maintenance might be an architectural response to continuous learning under finite capacity. Whether that hypothesis holds remains a question for testing, not a demonstrated need for machines to sleep like animals.