Meta Pairs an Agent With a Second One as Memory Coach

Why does an AI agent keep repeating a command it already tried and failed? Because on a long task it forgets what it learned three steps ago. Meta's answer, published this week, is almost organizational: give the agent a coach. A second agent that does no work of its own, watches the first, and steps in only to remind it what it already knows.

How the coach works

The setup pairs an unchanged "action agent" with a separate "memory agent." The memory agent reviews recent steps at fixed intervals and decides whether to inject a targeted reminder or stay silent. It keeps a structured memory bank in three parts: a private status field tracking progress and risks that the action agent never sees; a knowledge memory for stable facts like requirements, file paths, and configurations; and a procedural memory recording attempted commands, outcomes, and rejected hypotheses. Updates happen only through predefined tool calls, which keeps the bank from filling with noise.

The numbers

On Terminal-Bench 2.0, first-attempt task completion rose from 38% to 46%, an 8-point gain. On Tau2-Bench, the task-weighted average went from 55% to 62%, up 7 points. The gains weren't uniform: airline and retail tasks improved about 10 points each, while telecom moved only 3. In Meta's tests the memory agent ran on Claude Opus 4.6 while the action agent used the weaker Claude Sonnet 4.5, a deliberate pairing that suggests the coaching role benefits from a stronger model even when the worker is cheaper.

Why this matters beyond the benchmark

The interesting signal isn't the score, it's the architecture. Instead of waiting for one bigger model to remember everything, Meta split memory into its own role. That's a pattern a business can borrow long before it matters at frontier scale: the reliability problem in long agent tasks is often not raw capability but context that leaks away mid-run. If you're piloting agents on multi-step work, document paths, failed attempts, and rejected approaches, and feed that back in mid-task, the improvement here came from structure, not from a smarter core model. The paper and code are public, so this isn't a locked capability; it's a design choice available to anyone building on agents today.

Sources: The Decoder, arXiv