Free guide

Claude Dreaming: How Claude Fixes Its Own Memory (and How to Try It)

Anthropic's new post on agent memory, in plain English: why saving everything backfires, how a dreaming pass fixes bad notes, and what you can set up today.

By Beau Johnson·
The Short this guide goes with: Claude Dreaming: How Claude Fixes Its Own Memory (and How to Try It)

Quick answer

Dreaming is a background pass where Claude rereads past session transcripts, checks its saved memory notes against what really happened, fixes wrong notes and reorganizes the rest. Anthropic described it in its Nov 6, 2026 Engineering post "The components of agent memory". Developers can request access to Dreams, a research preview in Claude Managed Agents, and anyone using Claude Code can review its auto memory today with /memory.

The short version

  • In Anthropic's Pokémon test, Claude wrote one wrong note about where it was, then dropped back onto the route it came from three times. After a dreaming pass corrected the note, it went back into the cave and reached the next city.
  • Saving more isn't better memory: Claude Sonnet 3.7 wrote 413 memory files and earned one badge, while Claude Fable 5 kept nine files plus one main file (global.md) and earned four.
  • Dreams is a research preview in Claude Managed Agents. Claude Code's auto memory, which you can check today, keeps a short index (MEMORY.md) and loads only its first 200 lines or 25 KB each session.

How it actually works

A memory note is a prediction

When an agent writes a note, it's betting that the note will matter later. Anthropic says frontier Claude models have gotten much better at that bet, but no model really knows what future sessions will need, so some notes are wrong from the start and some go stale.

Saving everything backfires

In Pokémon Red, Claude Sonnet 3.7 saved a separate file for almost every battle, item and character: 413 files across 225 folders by about step 10,000, and one badge. The game changes all the time, so that archive kept going out of date. Claude Fable 5 saved only what the screen doesn't show (goals, plans, tested rules) in nine files plus a global.md dashboard, and earned four badges. Different models, one run each, so the lesson is judging what is worth saving, not saving less for its own sake.

Dreaming is a night-shift editor

Think of an editor who comes in after everyone goes home and checks the day's notes against the recordings. Dreaming does that for an agent: it reads the original transcripts, which are the ground truth, next to the memory notes, then fixes mistakes, merges duplicates and builds things like linked wiki pages that are easier to search. In the Pokémon run, it caught that Claude had written down the wrong side of Mt. Moon and rewrote the note.

It's not free

Dreaming uses compute, and Anthropic says having Claude constantly recheck every memory might not be practical. Its suggestion is to focus dreaming on information that's likely to change, or tune how often it runs. On one legal task the wiki even grew larger than the files it was built from.

How to set it up

  1. 1.Open Claude Code's memory panel

    In any Claude Code session, run /memory. You can browse and open memory files there and see the auto memory toggle. Auto memory is on by default in local sessions.

    /memory
  2. 2.Do your own dreaming pass on the notes

    Auto memory is plain markdown in your project's memory folder: a MEMORY.md index plus one file per memory. Open it, delete notes that are wrong or stale, and keep the index short, because only the first 200 lines or 25 KB of MEMORY.md load at the start of each session.

    ls ~/.claude/projects/*/memory/
    open ~/.claude/projects/<project>/memory/MEMORY.md
  3. 3.Developers: request access to Dreams

    Dreams is a research preview in Claude Managed Agents. On the Dreams docs page, click Request access. A dream reads an existing memory store plus 1 to 100 past sessions and writes a new, reorganized store. Your original store is never changed, so you can review the result and throw it away.

  4. 4.Run your first dream

    With access, create a dream from Python (the SDK sets the beta headers). It runs in the background for minutes to a few hours; poll it until it's completed, then attach the output store to your next session.

    dream = client.beta.dreams.create(
        inputs=[
            {"type": "memory_store", "memory_store_id": store_id},
            {"type": "sessions", "session_ids": [session_a, session_b]},
        ],
        model="claude-opus-5-5",
        instructions="Focus on coding-style preferences; ignore one-off debugging notes.",
    )
  5. 5.Or build your own dream

    Anthropic's cookbook recipe "Dreaming with a dynamic workflow" builds a small dream you control (prompts, models, which transcripts) and can run every night. Its example run over eleven sample transcripts cost $0.36 at list price and took about two and a half minutes. It needs Managed Agents beta access.

The fine print

  • Dreams is a research preview: you have to request access, default rate limits apply, and only some models are supported.
  • Dreams are billed at standard API token rates for the model you pick, and the cost grows with the number and length of sessions. Start with a small batch.
  • Managed Agents sessions, including dreams, are not eligible for zero data retention or HIPAA.
  • Claude Code's auto memory stays on your machine. It isn't shared across machines or cloud environments.
  • The Pokémon and wiki numbers come from Anthropic's own tests in its Nov 6, 2026 post, not an independent benchmark.

FAQ

What is Claude dreaming?

Dreaming is a background pass where Claude rereads past session transcripts and its memory notes, fixes notes that turned out to be wrong, and reorganizes its memory so future sessions can find things faster. The idea builds on sleep-time compute, proposed by researchers at Letta and UC Berkeley.

Can I use dreaming in the Claude app?

Not as a button. Dreams is a research preview for developers in Claude Managed Agents, and you have to request access. In Claude Code you can review and clean up auto memory yourself with /memory.

Why didn't saving more files help Claude remember better?

In Anthropic's Pokémon test, Claude Sonnet 3.7 saved 413 files that kept going out of date, and at one point spent 17 steps creating and deleting memory files without playing. Claude Fable 5 saved only plans and tested rules in nine files plus one main file (global.md) and got much further.

How much of Claude Code's memory loads each session?

The first 200 lines or 25 KB of MEMORY.md, whichever comes first. Claude keeps that file as a short index and moves details into separate topic files it reads when needed.

Does a dream change my existing memory?

No. A dream writes a new memory store and never modifies its inputs. You review the new store, then use it or delete it.

Sources

Facts checked 2026-11-07 against Anthropic's Engineering post (Nov 6, 2026), the Managed Agents Dreams docs, the Claude Code memory docs and the dreaming cookbook.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides