Meta AI proposes a second AI agent to keep long tasks on track

In their paper, Meta AI researchers describe a recurring problem where AI agents lose track of requirements or repeat failed actions over long tasks, a phenomenon they call 'behavioral state decay'. They propose a plug-and-play memory module that runs a separate 'memory agent' alongside an unchanged 'action agent', observing recent steps at fixed intervals, updating a structured memory bank, and deciding whether to inject a concise, memory-based reminder into the next action call or to stay silent. The memory bank is divided into private status, knowledge, and procedural memories, and the agent manages the bank through predefined tool calls rather than free rewriting. Tests were conducted on Terminal-Bench 2.0 and τ2-Bench with Claude Opus 4.6 as the memory agent. With Claude Sonnet 4.5 as the action agent, the system solved 46% of Terminal-Bench tasks on the first attempt, up from 38%, and improved the τ2-Bench task-weighted average from 55% to 62%, with notable domain differences: Airline and Retail improved by about 10 percentage points each, while Telecom only improved by 3. The stronger Opus 4.6 also improved but to a lesser extent, suggesting the benefit goes beyond compensating for limited capacity. Ablations showed that selective intervention and a maintained memory bank are key, and the approach outperformed the productive memory layer Mem0. The team also explored training an open model, Qwen3.5-27B, as a memory agent, finding that supervised fine-tuning and reinforcement learning were needed to properly calibrate interventions. Meta AI published the code on GitHub.