Anthropic Explains Agent Memory: 'Dreaming' Helps Claude Remember

What happened

The post defines agent memory as knowledge Claude gains from experience and retains between sessions, saved as context it can read later; it excludes changes to model weights, in-context learning within a single conversation and compaction summaries of older turns. The division of labor borrows from the human brain: the hippocampus captures experiences quickly, then replays and distills them into long-term knowledge while offline. Agents, beyond logs and transcripts, use idle compute to reflect, fix mistakes and write new notes and representations — a process the team calls dreaming, similar to the sleep-time compute proposed by Letta and UC Berkeley.

The evaluations show memory quality improving across model generations. On the Continual Learning Bench (CL-Bench) 1.0, Claude Opus 5 was the first model to close the learning loop: after a wrong answer it recorded the failure, proposed theories, tested them and wrote the findings into a rule; Claude Sonnet 5, Opus 5.5 and Sonnet 5.5 did the same. In a Pokémon Red test, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files in two folders plus a single global.md and earned four badges, and Opus 5.5 beat the game with a small folder of memory notes. Fixed rules for what to write to memory weakened stronger models; letting them judge salience freely worked better. Prompting Sonnet 4.6 to read its memory files before answering and then test its theories nearly doubled its benchmark score, while Sonnet 5 did this on its own 9 out of 10 times.

Dreaming also corrects wrong memories: in the Pokémon test Claude had written down the wrong side of a mountain for the PokéCenter, and dreaming checked the note against the session transcript, caught the error and rewrote it, after which Claude went straight to its goal; the run with the uncorrected note dropped back onto the old route three times and gave up. Dreaming also builds wiki-style knowledge maps: on a corpus of about 350,000 internal files, roughly 420M tokens, scores on 25 questions grew from 13% to 48% over 128 dreaming sessions, and with search plus the wiki the aggregate score rose from 54% to 61%, saving about 20,000 tokens per question. On a contract-terms task, an improved wiki raised the facts covered in a generated brief from 41 of 57 to all of them. The post acknowledges memory can still be wrong and drift, and continuously checking all memories is not practical.

Key facts

1. Published by: Anthropic, research post "The Components of Agent Memory," Nov 6, 2026, updated 21:15 the same day
2. Definition: agent memory is knowledge Claude gains from experience and retains between sessions, saved as readable context; it excludes weight changes, single-session learning and compaction summaries
3. Dreaming: idle compute organizes memory; sleep-time compute from Letta and UC Berkeley is a similar idea
4. CL-Bench 1.0: Opus 5 was the first to close the learning loop; Sonnet 5, Opus 5.5 and Sonnet 5.5 did the same
5. Pokémon Red: Sonnet 3.7 wrote 413 files in 225 folders and earned one badge; Fable 5 wrote nine files and earned four badges; Opus 5.5 beat the game
6. Rules vs freedom: fixed memory rules weakened stronger models; free judgment of salience worked better
7. Prompting effect: prompting Sonnet 4.6 to read memory files and test theories nearly doubled its benchmark score; Sonnet 5 did this on its own 9 out of 10 times
8. Dreaming correction: the wrong note was rewritten against the transcript; the uncorrected run dropped back three times and gave up
9. Wiki maps: on 350,000 files and about 420M tokens, 128 dreaming sessions raised 25 questions from 13% to 48%; search plus wiki rose from 54% to 61%, saving about 20,000 tokens per question
10. Limits: on a contract task the improved wiki raised brief coverage from 41 of 57 facts to all of them; memory can still be wrong and drift, and checking all memories continuously is not practical