Claude Managed Agents Dreams (Dreams for short, and you can think of it as a “dreaming” feature) is Anthropic’s tool for tidying an AI agent’s memory. An AI agent is an AI that can take actions on its own. One thing up front: Managed Agents is a feature developers use by writing code (through an API, which is a way for programs to talk to each other), not a button in the claude.ai chat window.
What Claude Dreams is
Picture an assistant who jots down key points in a notebook every day. A month later, the same thing has been written down three times, two pages contradict each other, and one page lists a phone number that was replaced long ago. Finding anything takes a lot of flipping.
Dreams is like recopying the whole notebook into a clean version over the weekend. In the official docs’ words, Dreams reads the existing memory store and past session transcripts and produces a reorganized new memory store. Duplicates are merged, outdated or overturned entries are replaced with the latest, and newly noticed patterns are added.
Two terms first:
- Memory store: a set of text documents that Claude can read and write. The agent uses it to carry information into its next job.
- Session transcript: the complete record the agent left behind from each past job.
Why Claude memory needs tidying
The docs explain that each Managed Agents job starts from a blank slate by default, and any state built up disappears when the job ends. A memory store lets the agent keep user preferences, project conventions, past mistakes, and background knowledge across jobs.
The problem is how the memory gets written. The agent writes memories while it works, focused only on the task in front of it, so the notes come out scattered and partial. After enough jobs, the memory store piles up duplicates, contradictions, and outdated entries. If the same preference is recorded twice, or an old preference and a new one sit side by side, the agent does not know which to trust. So it is not enough for Claude memory to just store things. It also needs regular tidying.
How Claude Managed Agents Dreams works
One Dream is an asynchronous job: you send it off and it runs in the background, so you do not have to wait. It needs two things:
- An existing memory store
- Session transcripts from 1 to 100 past jobs
When it finishes, it produces a separate new memory store. The docs say it usually takes minutes to hours, depending on how many transcripts there are.
The original memory store is not touched. You can inspect the new one. If you like it, connect it to future jobs. If not, delete it or archive it.
Before it runs, you can also attach a note on how to tidy, for example “keep only long-term work habits and ignore temporary chores.” The docs warn, though, that Dreams does an overall summary, not line-by-line editing. A request that points at one specific line, such as “reword sentence 3,” usually has no effect. To change a single memory, edit that entry directly in the new memory store.
Can you use Claude Dreams now
It is currently a research preview (an early test version), and you have to apply through the official form to use it. The docs list the Claude API as the supported surface, and only some models are supported, including Claude Opus 5.5, Claude Sonnet 5.5, and Claude Fable 5.1. The memory store that Dreams depends on is also still in beta.
If you only chat with Claude on the web or in the app, the docs describe use through the API, so for now treat this article as a look at where Claude memory is heading.
To understand how AI memory works first, see How AI Agent Memory Works. More Claude features are collected on the Claude hub.
Further reading
For whether Dreams costs extra and how long data is kept, see the FAQ above.
Cover image source: Anthropic (claude.com/blog/new-in-claude-managed-agents)