Skip to Content

Cleaning up Claude's memory: what the Dream feature does

An AI memory quietly goes stale. Anthropic lets Claude tidy it up itself, and you can rebuild that.
November 18, 2026

Between conversations, Claude remembers what it learns about you: your preferences, your role, fixed facts. That memory quietly goes stale. Old notes stay put, duplicates contradict each other, and your answers slowly get worse without you seeing why. That is why Anthropic is testing Dreams: Claude rereads earlier sessions and tidies up its own memory.

What happened

On 17 November, BenAI made a video about it and rebuilt it as a skill. We set his account next to what Anthropic itself publishes.

What Anthropic itself says

  • Dreams is a feature of Claude Managed Agents, for those building their own agents through the API. It is in research preview; you request access.
  • A dream reads an existing memory plus 1 to 100 earlier sessions, and produces a new, cleaned-up memory: duplicate notes merged, outdated or conflicting ones replaced by the latest state.
  • The old memory is never changed. You review the result and discard it if you do not like it.
  • In Anthropic research, Claude with dreaming grew a knowledge base over a large collection of documents. From two to 128 sessions, the score on a fixed question set rose from 13 to 48 percent.

What the video adds

  • BenAI shows where the memory lives in Claude Code: a general CLAUDE.md, a CLAUDE.md per project folder and an index with separate topic files that Claude maintains itself.
  • His free skill is a rebuild of Dreams that you can run every night. We could not verify the figures he gives about his own results.

Our view: a second brain only works if someone maintains it. Let AI propose the cleanup, but approve the changes yourself.

What this means for your organisation

If you work with Claude memory, a second brain or fixed instruction files, cleaning up is part of maintenance, just like tidying a shared drive. Anthropic itself warns that a longer instruction file is followed less well. Remembering more is therefore not automatically better.

What you do now

  • Ask Claude where its memory lives and have it show the contents. Remove what is no longer correct.
  • Keep your main instructions short. Put details in separate files that Claude refers to.
  • Do not run a cleanup skill from a video unread on your own conversations. Read it first and only let changes through after your approval.
  • Schedule a fixed moment, monthly for example, to go through your memory and instructions.

How it works under the hood

A dream is an asynchronous job that takes an existing memory store and a list of session transcripts as input. It runs from minutes to a few hours, depending on the number of transcripts, and delivers a separate new store. A short instruction steers the focus, for example "pay attention to writing preferences"; targeted one-line corrections you make yourself afterwards on the new store. According to BenAI, his rebuild uses subagents that compare the memory with the transcripts per group of conversations, and holds back doubtful cases for review.

This news touches chapters 10 and 15 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.

What you can do with this

Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.

Written in response to publications by BenAI (YouTube) and Anthropic (research on agent memory).