Anthropic's "dreaming" feature for Claude Managed Agents runs memory consolidation during idle/sleep periods — modeled on the mammalian brain's sleep algorithm. The three-step process: 1. **Replay** — reviews recent interaction sequences 2. **Pattern extraction** — identifies what mattered, what generalizes, which sequences predicted reward 3. **Curate/prune** — removes low-signal memories, reorganizes the trace from short-term storage into long-term schema This is qualitatively different from standard agent memory (vector search over embedded sessions). Vector retrieval works for facts. It does not extract patterns and does not reorganize the trace — which is why retrieval-only agents plateau: memory volume grows while real capability flatlines. The neuroscience basis: György Buzsáki's sharp-wave ripples. During sleep, the hippocampus replays recent experience at 10-20x speed, triggering corresponding reactivations in the cortex. The replay is causal — disrupt sharp-wave ripples with optogenetics and the rat fails the next day's task. Dreaming as Anthropic implemented it maps to: - Replay → session review - Pattern extraction → schema update - Pruning → curation of low-relevance memories The implication: "more context window" was always a partial answer. A 2 million token context is a hippocampus that never sleeps — it holds more, retrieves faster, and degrades in exactly the way Walker measures: signal-to-noise drops and encoding for new content weakens. ## Source - [[Anthropic just shipped sleep into agents]] — Aakash Gupta (@aakashgupta), X, 2026-05-07 — https://x.com/aakashgupta/status/2052275497032462392/?s=12&rw_tt_thread=True