Devin gets persistent memory and nightly 'dreaming' to prune it
Cognition keeps Devin's cross-session lessons in a Git drive and open-sources the format; plus an open 27B decision model you can self-host for agents.
Copy markdown
Memory lives in a Git repo you can read and edit
Devin now writes what it learns about your project — your package manager, naming conventions, deploy quirks, past environment failures — as short notes in a per-user 'Memory Drive' Git repo you can inspect, edit, and diff. It's live at no extra cost under Customize → Memory on devin.ai.
'Dreaming' recompacts that memory every night
A daily asynchronous pass consolidates overlapping notes, strips stale or one-off details, and back-fills lessons the live session missed — keeping the index small and signal-dense instead of ballooning over time. Cognition also open-sourced the Agent Memory Repo standard so other agents can read and write the same format.
AutoTrust's JEV-27B: a self-hostable decision model
Apache-2.0 on Hugging Face, JEV-27B bolts a decision head onto Qwen so it scores every option — text or image — in a single forward pass via a POST /v1/decide endpoint on vLLM, with no token-by-token generation. AutoTrust reports 78.3% on VL-RewardBench (ahead of GPT-5 and Gemini-3-Flash) and 95% success over 60 multi-step browser tasks, making it a cheap reward or routing brain for your own agents.