Anthropic agent memory research: dreaming fixes bad notes

Anthropic published its agent memory research on November 6, and it is the most practical document yet on how an AI agent should remember things between sessions. The post, The components of agent memory, shows that memory helps a lot. It also shows that a wrong memory hurts. The fix it proposes is "dreaming": the agent uses idle time to check its own notes against what actually happened.

What actually happened

Anthropic defines agent memory as knowledge an agent gains from experience and keeps between sessions as readable context. Weight changes and compaction do not count. The post splits it into four layers:

• Transcripts: a searchable log of what happened in each session.
• Memory notes: files the agent writes about important events.
• Representations: structures like a wiki of linked pages that the agent maintains.
• Injected context: the small file the harness loads into every session. A wrong note here goes into every run.

Dreaming runs between sessions. It corrects notes against the original transcripts, builds the wiki, and moves only the most useful notes up into injected context.

The numbers are large. On the CL-Bench 1.0 database task, run on Claude Managed Agents with a persistent memory folder, Sonnet 5.5 went from 0.644 without memory to 0.833 with it. Opus 5.5 went from 0.655 to 0.838. Older models gained less: Sonnet 4.6 went from 0.330 to 0.364. A better instruction ("read your memory first, test your theories") raised it to 0.641.

On a 420M-token code corpus, a wiki the agent built by dreaming rose from 13% to 48% of graded criteria as dreaming scaled. Wiki plus search scored 61% against 53% for search alone, and saved about 20,000 tokens per question. Anthropic also lists the limits: dreaming costs compute, wikis can collect errors, and on one legal task a wiki gave no clear gain.

Why it matters for your business

If you run an agent for support, ops or bookkeeping, it is probably already writing notes. In Claude Code, that is a MEMORY.md file, and the post says Claude Code loads only its first 200 lines or 25 KB. Nobody checks most of these files. One bad note ("refunds over $50 need approval" when the rule is $100) becomes a standing instruction.

Three things to do this week:

• Read your agent's injected context. Delete anything stale or unverified.
• Keep it short. Move details into linked files. Leave only the index at the top.
• Schedule a review job. A nightly job that checks notes against the logs is the cheap version of dreaming.

Sources: Anthropic, The components of agent memory.