Between sessions, Claude can correct errors in its memory

“Dreaming” lets Claude agents review past sessions, correct their notes, and build wikis. Anthropic explains how it tested the approach.

At the foot of Mt. Moon in Pokémon Red, Claude thinks it has already made it through the cave. A note places the Pokémon Center on the wrong side of the mountain. Following that note, the agent steps off a ledge and lands back on the route it came from. It makes the same mistake three times and ends the run back in the previous town.

In another run described by Anthropic’s engineering team, Claude reviews the gameplay log between sessions, spots the mistake, and rewrites the note. When play resumes, it heads back into the cave and eventually reaches Cerulean City. “Dreaming” is Anthropic’s term for this pass between sessions, when the agent revisits what it recorded while working. Checking the record before passing a note forward

In Anthropic’s experiments, memory takes the form of files that later sessions can consult. An agent might record a rule it has verified or a hypothesis it still needs to test. During a dreaming pass, the agent can check those notes against session logs, correct errors, merge duplicates, or group information by topic. The idea builds on research by Letta and UC Berkeley into sleep-time compute, which explores using the time between interactions to prepare for future requests.

The Dreams feature in Claude Managed Agents applies this method to a set of sessions chosen by the developer. By default, it creates a new version of the memory store that can be reviewed before it’s attached to future sessions. The API feature is available by request as a research preview. A wiki alongside file search

Opus 5 agents built a wiki over several passes from a collection of roughly 350,000 internal Stainless files. They created pages, checked information against the source documents, and linked related notes. Across the 25 questions in Anthropic’s internal evaluation, file search alone met 53% of the grading criteria. Combining search with the most developed version of the wiki raised the score to 61%. The wiki alone reached 48% at best.

Anthropic also reports that using both tools saved about 20,000 tokens per question during the answering phase. The full cost also includes the computation needed to build and update the wiki. In a separate exercise where a search of a single document often turned up the needed information, the wiki offered no clear improvement. One page per contract keeps agents from overwriting each other

In a simulated acquisition, several agents summarized contract clauses in the same table. Some entries disappeared when one agent saved its version over another’s. The legal brief produced from that first wiki included 41 of the 57 required facts.

The team then gave each contract its own page and assigned one agent to write it, with instructions to quote the clauses verbatim. In that exercise, the brief produced from the revised wiki met all 57 grading criteria.