Anthropic Says Offline 'Dreaming' Lets Claude Correct and Organise Its Own Memory

The post, written by Lance Martin and published on Anthropic's engineering blog on November 6, reports internal experiments rather than a product launch. It is relevant to any company building AI agents, because memory is one of the main obstacles to running software agents on long tasks that span many sessions. An agent that forgets what it learned, or keeps acting on a wrong note, repeats its mistakes. Anthropic's results suggest newer models are better at judging what is worth remembering, and that a background review can catch errors before they compound.

The idea borrows from neuroscience. The authors compare it with the way the hippocampus replays experiences during sleep so they can be consolidated into long-term memory, and say the work builds on the "sleep-time compute" concept from Letta and UC Berkeley. Anthropic splits agent memory into four layers: transcripts that act as a searchable event log, notes Claude writes during a session, larger structures such as wikis, and a small block of context loaded into every session. Changes to model weights are outside its definition.

Newer Models Close the Loop

On a database question-answering task called CL-Bench, where each question started in a fresh session with access to a shared memory folder, memory lifted Claude Opus 5.5's mean reward from 0.655 to 0.838 and Opus 5's from 0.564 to 0.828. The older Sonnet 4.6 barely improved, moving from 0.330 to 0.364. The authors wrote that it generated theories without testing them. When prompted to read its memory and check those theories, its score rose to 0.641.

Tight rules about what to save backfired. Telling models to record only their failures cut Opus 5's memory gain by 0.234 and Opus 4.7's by 0.150, which the authors read as evidence that frontier models do better with light guidance than with fixed instructions.

A long-running game of Pokemon Red showed the same shift. Claude Sonnet 3.7 saved almost everything, writing 413 files across 225 folders totalling about 977,930 characters, and earned one badge. Opus 5.5 kept a small set of notes and finished the game. In a separate run, dreaming corrected a note that had placed a landmark on the wrong side of a cave. With the fix, Claude reached its next destination at action 1,935. A run with no memory had still not got there after 3,000 actions.

Wikis Help Most Alongside Search

The 13%-to-48% result came from a corpus of roughly 420 million tokens from the software company Stainless, tested against 25 questions. The wiki on its own still trailed plain search, which scored 53%. Used together, search and the wiki rose from 54% to 61% as the wiki grew, and saved about 20,000 tokens per question. Over the run the wiki grew from 103 pages with 766 links to 222 pages with 12,893 links.

On 24 merger questions from Harvey's legal benchmark, Claude scored 87% with a wiki against 73% with search alone. The authors reported no clear gain where answers could be found by text search in a single document, and one wiki ended up 1.77 times the size of the material it summarised. When several agents wrote to one shared table, 41 of 57 contract facts survived. After the instructions were changed to give each contract its own page with verbatim terms, all 57 did.

The authors cautioned that the layer loaded into every session must stay small, because wrong notes spread and long instruction files reduce how closely models follow them. Claude Code's existing auto memory, they noted, loads only the first 200 lines or 25 KB of its memory file. All of the results come from Anthropic's own evaluations. The company listed when to write memory, how to keep it current and how often agents should dream as open questions, and has published starter code for dreaming and knowledge wikis on its developer platform and GitHub.