Meta gives AI agents a coach whose only job is reminding them what they already know
Meta AI researchers name a failure mode they call 'behavioural state decay': in a long task, the requirement or the failed command is still sitting in the agent's context window, but by the time it has to act, that information no longer shapes the decision. Their fix leaves the working agent untouched and runs a second, smaller 'memory agent' alongside it, keeping a structured bank of status, facts and procedures and deciding at every step whether to inject one short reminder or stay silent. On the Terminal-Bench 2.0 suite, Claude Sonnet 4.5 rose from 37.6% to 45.9% pass@1 — 8.3 points — and on τ²-Bench from 55.0% to 61.8%. The code is public.
Why it mattersNearly every other attempt to fix agent memory adds more retrieval: pull more text into the window and hope the model reads it. This paper argues the problem is the opposite — agents are drowning in their own transcripts, and what they need is an editor, not an archive. If a small supervising model can buy eight points on a hard benchmark without touching the main one, that is an unusually cheap upgrade for any long-running coding or support agent.
▶ Related video: Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents (Jul 2026)
Read in the app — free, in 9 languages
Related stories
Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down2026-08-21Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21