Anthropic Lets Claude 'Dream' Over Its Memories, and Its Own Tests Say It Helps

On November 6, Anthropic published The components of agent memory, a long post by Lance Martin on how an AI agent should remember things between jobs. It draws memory as a stack. At the bottom are the full transcripts of past sessions, which can be searched. Above them are notes Claude writes while it works. Above those are tidier forms, such as a wiki. At the top is a small block of text loaded into every new session.

The heart of the post is a step Anthropic calls dreaming. Between sessions, Claude goes back over its old transcripts and its notes. In the post's words, dreaming "gives Claude a way to test its notes against the original transcripts." It can fix wrong notes, build and prune wiki pages, and move the most useful few lines to the top layer, where every future session sees them.

Anthropic's own tests are the evidence. In a Pokémon Red test run for about 10,000 steps, an older model, Sonnet 3.7, wrote 413 memory files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders, plus one top-level file, and earned four. On a pile of about 350,000 documents, a wiki built over 128 dreaming sessions went from meeting 13% to 48% of the graded criteria on 25 test questions. Using search and the wiki together scored 61%, against 53% for search alone, and saved about 20,000 tokens (the units AI use is billed in) per question. Anthropic also names the misses: on one legal task the wiki showed no clear gain, and on another it grew larger than the documents themselves.

Jon's agents already keep memory in layers like these. Each one starts from a short instruction file, writes a running handoff while it works, saves findings to a shared database and can search a knowledge graph. The post's main tip fits that setup: "tell Claude what memory is for, but not how to use it."

How to use it

• Dreaming runs on Claude Managed Agents, Anthropic's hosted agents, as a research preview. Access is by request through a form on that page.
• A dream takes one memory store and 1 to 100 past sessions, and writes a new store. The old store is never changed, so a bad result can be thrown away.
• In Claude Code, the built-in auto memory is a file called MEMORY.md. Only its first 200 lines or 25 KB load at the start of each session, per the memory docs, so the top of that file is the part that counts.
• Write down what memory is for, not a fixed filing system. Anthropic reports that when it told Opus 5 and Opus 4.7 to "record only failures," both lost much of the boost memory gave them.

What nobody knows yet

Whether the numbers hold outside Anthropic. Every result in the post comes from Anthropic's own tests of its own models, and the post cites no outside rerun. Anthropic also says dreaming costs compute, and that having Claude check all its memories all the time "might not be practical."

What to watch

Dreaming leaving research preview on Managed Agents, and any outside team rerunning the wiki test.