Claude can now “dream”

Designed by Magnific

What does this mean?

Before memory, most AI models including Claude were stateless, that is, each of their sessions were independent. The model receives a prompt, generates a response and the connection closes.

So everytime an agent started a new session, it started completely fresh! No memory of what worked last time, what failed and what patterns it had spotted. And then came memory, it helped but got worse over time, as after multiple sessions, auto-memory accumulates noise like stale facts, contradictions, debugging notes about files that no longer exist.

Wait, is dreaming different from memory?

Yes. Memory is the storage. It is what you told Claude, what it learned and what happened in past sessions. It just accumulates over time.

Dreaming is the maintenance of this storage. It goes through the memory periodically and not only removes noise but also surfaces useful patterns

Without Dreaming, memory just grows. And growing memory isn’t always good, after multiple sessions, it becomes filled with old, contradictory, outdated information. The agent technically remembers everything but is now confused. So the memory that was supposed to help Claude remember instead became clutter.

Dreaming solves that. It doesn’t give the agent more memory, it gives it better memory. This helps agents self learn and iterate on their performance.

How does this work?

Claude Dreaming is a scheduled background process that reads an AI agent’s past session transcripts and existing memory store, then produces a new reorganized memory store, removing duplicates, replacing stale entries and surfacing new patterns. It is not a new model and it does not change Claude’s underlying weights. It is a maintenance layer that lives on top of an external memory store, so the agent gets better between sessions without retraining anything.

The key word is between, i.e, it runs when the agent is idle, not while it’s working!

It converts relative dates to absolute dates i.e, “yesterday we decided to use Redis” becomes “on 2026–05–21 we decided to use Redis” preventing confusion as memories age.

It deletes contradicted facts like if you switched from Express to Fastify three weeks ago, the old entry disappears.

It removes stale memories like debugging notes about deleted files.

It merges overlapping entries, if three sessions noted the same build quirk, they consolidate into one clean entry.

Image credits: Code with Claude 2026, San Francisco

Instead of requiring manual intervention, Claude can now inspect its previous sessions, identify skills it missed or lessons it should have learned and update its own memory accordingly. This way the agent can generate strategies like creating a “descent playbook” that improves future results without human intervention.

Results

One observed case showed that dreaming went through 913 sessions worth of accumulated memory and cleaned it up in 8 to 9 minutes, which could seem like a lot of time, but remember, it runs in the background without blocking your work!

Legal AI startup Harvey ran the pilot and saw task completion rates climb roughly six times after enabling Dreaming for legal-drafting workflows. This 6x jump in task completion wasn’t because Claude got smarter, it’s because it stopped wasting time on a messy, contradictory memory.

By default, dreaming runs automatically between sessions. But developers can also trigger it manually through the Claude developer console, to run simulations, where the agent can immediately combine learnings from those experiences and apply them to future tasks.

Source: Claude with Code 2026, San Francisco, Youtube-Julian Goldie SEO, Auto Memory and Auto Dream: how Claude Code learns and consolidates its memory-Antonio Cortés