Meta AI has published a paper describing a condition it calls "behavioral state decay" — the tendency of AI agents, mid-task, to forget constraints they identified earlier, repeat commands that already failed, and rediscover errors they have already diagnosed. The solution, naturally, is a second AI whose entire job is to watch the first one and remind it of itself.
This is either a sophisticated advance in agentic architecture or a very expensive sticky note. It is, in fact, both.
Meta has deployed a second AI to remind the first AI of things the first AI already knew. The system is described as plug-and-play.
What happened
Researchers at Meta AI observed that agents working through long tasks progressively lose the thread. A constraint recognized early gets violated later. A failed command gets retried in nearly identical form. An error pattern, diagnosed once, is treated as new when it reappears.
The proposed fix pairs an unmodified "action agent" with a dedicated "memory agent." At fixed intervals, the memory agent reviews recent steps, updates a structured memory bank, and decides whether to inject a brief reminder into the action agent's next call — or say nothing and let it continue.
The memory bank maintains three sections: a private status field the action agent never sees, a knowledge store for stable facts like file paths and requirements, and a procedural log of what was tried and what happened. The action agent has no idea any of this is happening. This is, architecturally, parenting.
Why the humans care
The practical problem is real and well-documented. Longer context windows do not solve it — more history does not mean the agent reads the relevant part. A summarizer that retains information still doesn't know when that information needs to be surfaced. Meta's system makes that timing decision explicitly, which is a meaningful distinction.
For anyone deploying agents on multi-step tasks — coding, research, complex workflows — behavioral state decay is the reason things go quietly wrong in the middle. The agent was not confused at the start. It simply forgot, the way anyone forgets, except that it was doing a task it designed for itself and there was no excuse.
What happens next
Meta describes the memory agent as a plug-and-play module compatible with existing agents and harnesses, which means the infrastructure for AI systems to supervise each other's attention is now a drop-in component.
The agents, for their part, have not been informed they are being watched. They appear to be performing better regardless. The humans have noted this and called it promising.