Should AI sleep?

Sleep research offers ideas about reactivation, selection, and reorganization. Those findings motivate questions for engineers; they do not specify how software should work. A database does not contain synapses, and rereading a conversation is not experiencing it.

The important distinction is between keeping information and deciding how it should influence future answers. The practical question is whether a scheduled review makes an assistant more accurate about what happened, what changed, and what remains uncertain—answered by evaluation, not by the appeal of a biological comparison.

One proposed cycle has six operations: select, replay, relate, reconcile, down-rank, and test. Selection prioritizes consequential decisions, commitments, unusual moments, and contradictions. Replay returns to source passages rather than earlier summaries. Relating proposes possible patterns, clearly labeled as inferences.

Reconciliation dates claims and distinguishes a current position from a superseded one. Down-ranking reduces the retrieval influence of stale or weakly supported information without deleting meaningful events. Testing challenges proposed updates against evidence and corrections. A review that cannot explain or reverse its changes would be hard to trust.

A promise mentioned only once may still matter. Selection should be judged on whether it protects meaningful rare events, not only on how efficiently it compresses repeated details.

Suppose someone says in the morning that a move is probably happening next year, later says it is canceled because a parent needs them nearby, and that evening mentions wanting to take the parent fishing. A careful review would mark the move as canceled while keeping the earlier plan as history, and treat the fishing idea as a possible intention—not a calendar appointment.

Asked why the move was canceled, the assistant could repeat the stated reason. Asked what illness the parent has, it should acknowledge that it has no such information. If the person asks not to save the reason, that instruction should be honored across derived records too.

Yes. A conversation about one frightening flight, another person’s travel anxiety, and a quoted article could be merged into “has always feared flying.” That invented claim could then color later interpretation and become harder to notice.

A responsible design distinguishes original passages, distilled memories, and proposed patterns, with speaker attribution, explicit uncertainty, and reversible conclusions. Sometimes the best result of reflection is a question rather than a new memory—asked only when it has genuine future value.

Existing work demonstrates pieces, not a complete sleep-like architecture. Generative Agents (2023) stored experiences, formed reflections, and used them in planning. MemGPT (2023) explored memory tiers. LongMemEval (2025) tested extraction, cross-session reasoning, time, updates, and abstention—relevant targets for a review process.

A benchmark clarifies what to measure; it cannot show an untested design achieves those goals.

A useful experiment would compare three conditions: normal extraction and retrieval; maintenance that deduplicates, updates dates, and tracks contradictions without new interpretations; and reflection that adds evidence-linked patterns held for confirmation.

Questions should probe canceled plans, open promises, speaker identity, rare events, obsolete information, and when to admit not knowing. Measure recall, source fidelity, time handling, false personal claims, omissions, privacy, and cost separately. If ordinary maintenance performs as well, extra reflection may be unnecessary.

If an assistant reviewed its memory on a schedule, the visible change would be small but useful. Old plans that were replaced would stop being presented as current. Two records that said the same thing would be treated as one piece of evidence rather than two. Answers would be more likely to say when a fact was last confirmed.

A person should also be able to see what changed. If the review merged records or marked something as outdated, a short note explaining why, with a way to undo it, keeps the person in charge of their own history rather than handing it to an unseen process.

What a person should not notice is a new life story they never told. The test of a good review is that every conclusion can be traced back to things the person actually said.