Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo:

In Agent Memory Repo, memory records live in files and are versioned with Git It’s a simple primitive but unlocks surprising use cases like a message board across your agent sessions

In early experimentation we saw a swarm of Devins using the memory as a way to coordinate at large scale without us prompting them to (we promise they didn’t hack anyone)

Read more about memory and dreaming at:

Finally I can retire my git brain repo

@cognition Tried it on the memory Claude Code kept for my project: 19 notes over 10 days. Made it a git repo. A fresh Devin session recalled what Claude Code had learned, then dreamed on it: 21 commits, 19 → 14 notes, duplicates merged, stale notes gone.

at least credit the progenitors of sleep-time compute.. cc @Letta_AI

What a name 😴 ✨

This is huge guys, you don't yet understand how good it can be on auto self improvement Especially SWE-2

Cool! You can benchmark it here at to see if it’s a fancy or efficient memory system. :)

Funny I just added my own "Night Shift" to my automations last week, to instruct my agents when I'm sleeping without intervention, gotta check this out.

Cool. Hope this improves the PR reviews. They've gotten very prone to flagging non-issues the last few days.

how do you measure memory effectiveness? how do you handle personal memories? for example, in the team working on same project, each member should have their own memory + shared ones this looks fine for small, single-developer repositories

Thank you for the OSS standard looks super useful

Devin built my studio brain repo like 3 days ago ... looks like I can now retire that, more devin work ! Having this working now plus devins being able to coordinate together is a massive unlock. I throw few devins in parallel working on few interconnected linear tasks so now they can work togehter and I can be less in the loop ! I love that and love that I can now just give my devins a mission and leave them handle the rest. This is clearly what @ScottWu46 mentioned in his interviews - what if agents could run for ..like a year and you just give a mission? Love that.

Memory of how you work is half of it. The other half is what you did outside the agent: the docs you read, the fixes you made by hand, the tab you left open at 6pm. That's the gap I've been building for with DayDream (open source, MCP). Excited to see this direction.

The edge case I'd test is a one-off instruction becoming a global preference. Memory needs scope: this repo, this task, or all projects. With two repos using conflicting conventions, does retrieval select the right preference before the agent plans?

Amazing way to do it, thank you guys for the open spec too!

This is huge. I use Devin a lot when im not at home and not explaining the same stuff every session would save me so much time

8 months later and its just @Letta_AI context repos with a few cron jobs sprinkled on top?

@Scobleizer Can it beat @ScottWu46 on codeforces yet?

How do you stop an older session from writing back a memory that dreaming deleted? I'd keep a tombstone for removed records; a missing file alone can't tell an agent "we deliberately forgot this."

Seems like a really good addition

@alexkaplan0 Great!

@Letta_AI original concept.

All the cool features ..coming in straight from Devin

Very cool! We've been using a version of this, but with one more step: There's the idea of personal, repo, and org-level memories, so it's a bit more of a hierarchical graph, but each level has self-improvement loops.

Everyone saying “message board” and “agents co-ordinating themselves” say it like these models are just acting with no supervision or some model directing the work. There has to be a driving force be it human or model. Having said that, a message board is primitive for communication between agents. Very inefficient

The Git-versioned memory files are the part I'd use. I've been hand-rolling session notes in text files for months.

does it hallucinate context from the old logs

🌚💤🌌

I like the “dreaming” direction. But there’s an interesting layer underneath memory consolidation: When a memory is created, merged, or removed, can the system preserve the evidence and history behind that change?

If I switch a project from SQLite to Postgres, does Dreaming catch the old notes that assumed a local file? Curious how you test those knock-on effects.

Cool. I've had something similarish? running on a qwen model on my mini for a few months. Nightly PCA identifies what's load bearing, compares to existing memory and edits the long and short term. If I don't talk about it then eventually the model 'forgets.'

Agent Memory Repo as OSS standard is huge move

Is this similar to the dream feature Anthropic was talking about releasing in their leaked documents earlier this year?

Devin learning your workflow while you sleep is genuinely wild progress