Claude Dreaming: How Anthropic Fixes Stale AI Agent Memory, and the Catch

An AI agent that keeps notes between sessions has a quiet problem: some of those notes are wrong, and some go out of date. Anthropic's post "The components of agent memory" (November 6, 2026) puts it this way: a note "that's wrong from the start or that drifts as the environment changes can be read, and trusted, by Claude in a later session." Its answer is a process it calls dreaming.

What dreaming is

Agents leave a record of their work in logs and transcripts. Dreaming means Claude goes back over those transcripts later, checks its memory notes against what actually happened, and rewrites them. Anthropic says dreaming can "correct mistakes, create notes, build new memory representations, or shape what's preemptively loaded into future sessions."

The clearest example in the post comes from Claude playing Pokémon Red. Claude wrote down the wrong side of a mountain for a landmark, then noted "Mt Moon DONE" when it wasn't. With that note it kept dropping back onto the route it came from, and the run ended without ever reaching the next city. Anthropic then ran dreaming on that memory folder: Claude checked the note against the session transcript, caught the error and rewrote it. With the fixed note, it went straight back into the cave and eventually reached Cerulean City.

On the Claude API, this is the Dreams feature of Managed Agents. A dream takes an existing memory store plus 1 to 100 past sessions and writes a new, reorganized store, with duplicates merged and stale or contradicted entries replaced. It never changes the input store, so you can review the output and throw it away. Jobs typically take minutes to a few hours, billed at standard API token rates for the model you select.

The numbers Anthropic reported

• A wiki over a big corpus. Anthropic had Claude build a wiki from about 350,000 internal files (roughly 420M tokens) from Stainless, a company it acquired in May. Over 128 dreaming sessions, the wiki-only score on 25 questions rose from 13% to 48% of graded criteria. With search plus the wiki, Claude scored 61%, versus 53% with search alone, and saved about 20,000 tokens per question.
• Merger questions. On merger questions that needed information from several documents, Claude scored 87% with the wiki versus 73% with search alone.
• Contract terms. When several agents wrote to one shared table, only 41 of 57 graded contract facts made it into the brief (72%). After Claude rewrote the wiki instructions so each contract got its own page, quoted word for word, all 57 made it (100%).
• Smaller memory, further progress. In one Pokémon run each, about 10,000 steps in, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders plus one global.md and earned four.

Anthropic also found that strong models do best when told what memory is for, and left to judge what to save. A fixed rule like "record only failures" cut much of the gain memory gave Opus 5 and Opus 4.7.

If you use Claude Code

Claude Code's auto memory works on the same idea of a small top layer. Per the Claude Code docs, "The first 200 lines of MEMORY.md, or the first 25KB, whichever comes first, are loaded at the start of every conversation. Content beyond that threshold is not loaded at session start." Topic files load only when Claude reads them. So keep MEMORY.md as an index, one line per memory, and put the detail in topic files.

The catch

• Dreams are a research preview. You request access, and the feature isn't ZDR or HIPAA eligible.
• These are Anthropic's own tests. Every number above comes from Anthropic's evaluations, with no independent replication cited.
• Wikis don't always help. On one legal-brief task where most answers sat in a single document, the wiki showed no clear gain over search. On another, the wiki grew larger than the documents it was built from.
• Dreaming costs compute. Anthropic notes that checking every memory all the time "might not be practical," so it suggests focusing dreams on information that's likely to change.

Here's how you get paid

My opinion, no income guaranteed: small firms sit on slow-changing piles of documents (contracts, filings, product docs) that people ask the same questions about. Anthropic's MIT-licensed knowledge wiki quickstart shows the pattern: read the corpus once into a wiki with a source on every fact, then answer from it. It says its own wiki costs about $19 to build once on its benchmark's average.

1. Build the wiki from their own documents. Start with one folder they actually query, like contracts or support docs.
2. Keep every fact tied to its source. That's what lets their team trust the answers and check them.
3. Charge a monthly fee to keep it fresh. New documents go in, stale facts get replaced, and you rerun the checks.

Dreams and Managed Agents memory stores aren't ZDR or HIPAA eligible, so check the client's data rules before any of their documents go in. Be upfront that the published results are Anthropic's own tests, and measure the wiki on their real questions before you promise anything.