Loading video...
Video Failed to Load
Today we shipped Memory and Dreaming in Devin: persistent self-cleaning memory for agent sessions. The standard is OSS and also works across any harness in any environment. ⚡
46,897 views • 6 days ago •via X (Twitter)
14 Comments

New documentation is up here:

Absolutely cooking brother. Cognition team are firing on all cylinders.

this is soo cool

awesome. working on adding dreaming to my agent app too

incredibly based🔥🔥 trying in few

very cool!! meanwhile

Agents that forget everything between sessions have been the quiet tax on using them. Devin cleaning its own memory overnight means it gets better at your job the longer you use it, not just the day you set it up.

Self-cleaning is the hard part. The failure to watch is a stale memory that still looks plausible, like an old build command or a renamed API. Tagging each memory with the session it came from and when it was last confirmed would make the pruning auditable.

@dabit3 does the self-cleaning keep session histories useful or does it drop context that agents need later? what's the retention policy look like?

Persistent memory makes agent workflows feel significantly more capable

Killing it

This is such a huge step for agentic workflows. We actually went deeper on this here:

How does dreaming distinguish stale notes from rarely needed ones? A rollback fix can sit unread for six months and still be the most valuable note in the repo.

persistent memory that cleans itself up is genuinely clever, nice ship
Related Videos
Sensitive content
LLMs are stateless. We built Dria Mem Agent to change that: Making memory a first-class feature. A 4B agent with local interoperable memory across Claude, ChatGPT and LM Studio. It turns LLMs from stateless chat into stateful agents with persistent human-readable memory.
Dria
170,822 views • 1 year ago

