How does Claude’s dreaming memory work, and how do you stop an AI agent’s notes from going wrong?

AI agents save notes between sessions, and a wrong note compounds: it is copied into the next plan and quoted back as fact. Anthropic’s Nov. 6, 2026 research describes “dreaming” — an offline pass that re-checks an agent’s notes against the original transcripts, corrects and merges them. Here is how it works, what Anthropic’s own tests showed, and five steps to keep any agent’s memory honest.

You ask an AI agent to finish a job, it picks up its own notes from last time — and walks back into the same dead end it hit yesterday.Short answer: AI agents like Claude save written notes between sessions so they can resume work, and Anthropic’s Nov. 6, 2026 research post, ‘The components of agent memory,’ describes a fix it calls “dreaming”: an offline pass that re-reads the agent’s notes against the original session transcripts, corrects errors, merges duplicates and deletes stale entries. In Anthropic’s telling, its agent playing Pokémon kept returning to a landmark because its notes insisted the landmark mattered — long after that had stopped being true.

The stale note, not the model, was doing the steering. Dreams is available in research preview in Claude Managed Agents.Why it matters: notes compound. A wrong note is copied into the next session’s plan, then quoted back as settled fact, and every later session builds on it.

In Anthropic’s own tests — company-run, not independent — accuracy on a wiki-style task rose from 13% to 48% as dreaming was scaled from 2 to 128 sessions. Treat the numbers as the maker’s own measurement; the direction is the lesson: unchecked memory rots, reviewed memory improves.Keep any agent’s notes honest, whatever tool you use:1. Keep the memory in one notes file you can open and read.

If you cannot see what the agent remembers, you cannot audit it.2. Review the file on a schedule — weekly is plenty — and scan for things that were true once: old prices, finished projects, superseded instructions.3. Delete stale entries outright instead of stacking corrections on top.

The agent reads the whole file, contradictions included.4. Verify important notes against the source record — the original email, invoice or transcript — never against the agent’s own summary of it.5. Date every entry, so you can tell fresh facts from fossils.Common mistake: letting the agent grade its own homework.

A note that merely sounds confident gets carried forward exactly like a checked one — the Pokémon landmark was wrong with total confidence.Ad’s take: this is the same discipline as cleaning out a toolbox: the tool is not broken, the drawer is just full of last year’s parts. Five minutes of pruning beats hours of wondering why the agent keeps doing the wrong thing.