Anthropic's Claude dreams between sessions — and fixes its own wrong memories

Anthropic published an engineering post on November 6 describing how Claude keeps what it learns between sessions — and the centerpiece is a process the lab calls “dreaming”: in idle time between sessions, Claude rereads its own logs, corrects notes that turned out to be wrong, and builds linked wiki pages from what it found.

The demos are disarmingly concrete. Playing Pokémon Red, one Claude wrote down the wrong location for a PokéCenter and walked off a one-way ledge three times before a dreaming pass caught the error against the transcript and rewrote the note. A leaner memory — ten files instead of 413 — earned four badges where the hoarder earned one.

On a bigger test, Claude spent 128 dreaming sessions turning about 350,000 internal documents into a wiki, lifting its score on 25 company questions from 13% to 48%. Anthropic is candid about the catch: dreaming costs compute, and a wrong memory injected into every future session can keep propagating — so how often to let Claude dream is still an open question.