Anthropic has published a detailed breakdown of how its Claude agents store and repair memory between sessions, including a process it calls "dreaming," in which an agent uses idle time to review its own logs and correct its notes. On the company's CL-Bench 1.0 database task, Claude Opus 5.5's mean reward rose from 0.655 without memory to 0.838 with a memory store.
The post, written by Anthropic's Lance Martin and published November 6 on the company's engineering blog, defines agent memory as knowledge Claude gains from experience and keeps between sessions as context it can read later. That definition excludes changes to model weights, in-context learning, and compaction, the summaries that replace older turns inside a context window.
Anthropic lays memory out as a four-layer hierarchy. Transcripts sit at the bottom as a searchable log of events. Above them are memory notes Claude writes itself, then richer representations such as wikis, and at the top a layer of injected context that loads into every session.
Key facts
- 01With a memory store, Claude Opus 5.5 scored 0.838 on CL-Bench 1.0's database task, up from 0.655 without memory.
- 02Anthropic's memory hierarchy has four layers, from searchable transcripts at the bottom to context injected into every session at the top.
- 03Opus 5.5 beat Pokemon Red with a small notes folder, while Sonnet 3.7 wrote 413 memory files and earned one badge.
- 04Across 128 dreaming sessions on a 420M-token corpus, Opus 5 agents raised graded criteria from 13% to 48%.
- 05The post, by Anthropic's Lance Martin, was published November 6, 2026.
Dreaming borrows from sleep-time compute, an idea the post credits to Letta and UC Berkeley, who proposed that models use the gaps between interactions to reorganize existing context. Anthropic applies it to memory: the agent reflects on its logs to fix mistakes or write new notes, and the original transcripts give it a ground truth to check those notes against. The post compares the process to the hippocampus replaying experiences during sleep.
The CL-Bench numbers show memory paying off more with each model generation. Sonnet 5.5 climbed from 0.644 to 0.833 with memory, Opus 5 from 0.564 to 0.828, and Opus 4.7 from 0.383 to 0.700. Sonnet 4.6 barely moved, from 0.330 to 0.364, because it read its memory before answering less than half the time, while Sonnet 5 did so 9 times out of 10. Telling Sonnet 4.6 to read its memory and test its theories lifted it to 0.641.
Anthropic also found that rigid note-taking rules can backfire on stronger models. An instruction to record only failures cut the memory gain by 0.234 for Opus 5 and 0.150 for Opus 4.7, against just 0.020 for Sonnet 4.6.
Pokemon Red shows the same trend. Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge. Claude Fable 5 kept nine files in two folders plus a global notes file and earned four badges. Opus 5.5 kept a small folder of goals and next steps and beat the game. In a separate Mt. Moon test, the run with dreaming-corrected memory reached Cerulean City at action 1,935, while the raw-memory and no-memory runs never got there.
At enterprise scale, Anthropic ran 128 dreaming sessions with Opus 5 agents over the codebase of Stainless, a software company it acquired in May: about 350,000 files and roughly 420 million tokens, tested with 25 questions. Graded criteria rose from 13% to 48% as sessions scaled from two to 128, and the agent-built wiki grew from 103 pages and 766 links to 222 pages and 12,893 links. Pairing the wiki with search reached as high as 61%, against 53% for search alone, and saved about 20,000 tokens per question.
On the Harvey LAB legal benchmark, a wiki raised accuracy on multi-document and judgment questions in a 24-question merger set to 87%, against 73% with search alone. On a contract-terms task, Claude went from 41 of 57 graded facts, or 72%, to 100% after agents built a separate page for each contract. On another LAB task, the wiki grew to 1.77 times the size of the source documents with no clear gain over search.
The post lists the limits plainly. Notes can be wrong from the start or drift as the environment changes, and Claude trusts them in later sessions. Dreaming costs compute, and injected memories use tokens and can carry a bad note into every session. In Claude Code, auto memory loads only the first 200 lines or 25 KB of the MEMORY.md file into each session.
Anthropic links the work to developer documentation for memory in Claude Managed Agents, its memory tool, and a guide to dreaming, though the post gives no pricing or release date for dreaming itself. The company closes by saying it still has "a lot to learn about designing Claude's memory systems."
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
