Meta AI uses dual-agent framework to improve memory

Meta AI researchers have designed a plug-and-play memory system that pairs an unmodified action agent with a secondary memory agent to combat what they term behavioral state decay during long-running tasks. In complex workflows, primary agents often lose track of constraints or repeat failed commands as their history gets buried in the context window. To solve this, the secondary memory agent periodically reviews recent steps, updates a structured memory bank containing private status, knowledge, and procedural records, and decides whether to inject a targeted reminder or remain silent.

The researchers evaluated the architecture using Claude Opus 4.6 as the memory agent across two key benchmarks: Terminal-Bench 2.0 and tau2-Bench. When paired with Claude Sonnet 4.5 as the action agent, the system boosted the first-attempt success rate on Terminal-Bench 2.0 from a 38 percent baseline to 46 percent. On tau2-Bench, the task-weighted average increased from 55 percent to 62 percent, with airline and retail domains seeing 10 percentage point jumps while telecom improved by 3 points. Even when using the stronger Claude Opus 4.6 as the action agent, performance rose by 2.4 percentage points on Terminal-Bench 2.0 and 2.5 points on tau2-Bench. The team also successfully trained a smaller Qwen3.5-27B model to act as the memory agent using supervised fine-tuning and reinforcement learning.

For AI practitioners, this framework offers a way to deploy more reliable autonomous agents without the high token costs, latency, and distractions associated with dumping entire transaction histories into a context window. By selectively intervening rather than constantly feeding the action agent the entire memory bank, the system outperformed alternative approaches like the Mem0 retrieval layer. This targeted intervention ensures that critical context, such as verifying tool data over user claims, is injected only when it actively shapes the next decision, making long-horizon agentic workflows far more viable in production.