Meta发现智能体做长任务会忘事，加了个外挂记忆模块，在Terminal-Bench 2.0上明显提升pass@1，而且即插即用。
Meta研究人员定义了智能体的“行为状态衰减”（behavioral state decay），即任务事实、先前尝试和未完成的子目标被埋没在上下文窗口中。他们提出一个即插即用的记忆模块：一个独立的记忆智能体维护结构化记忆库，每一步决定是否注入基于记忆的提醒。在Terminal-Bench 2.0和tau-squared-Bench上，该方法将pass@1提高了多个百分点，且不修改原有动作智能体。
New research from Meta. (bookmark it) It's on how to fix agents that forget previously made decisi...
New research from Meta. (bookmark it) It's on how to fix agents that forget previously made decisions. It's well know that long-horizon agents keep forgetting decisions they already made. Meta researchers give this failure a name, behavioral state decay, where task facts, prior attempts, and open subgoals get buried in the context window or pushed past it, so they stop influencing the next action. Their fix runs a separate memory agent alongside an unmodified action agent. It maintains a structured memory bank from the recent trajectory and decides, each step, whether to inject a memory-grounded reminder or stay silent. The module is plug-and-play with frontier agents and existing harnesses. It lifts pass @1 for both weaker and stronger action agents on Terminal-Bench 2.0 and tau-squared-Bench. Overall, they find that memory that actively surfaces the right fact at the right moment is a more useful primitive than passive retrieval that only fires when the agent thinks to ask. Paper: arxiv.org/abs/2607.08716 Learn to build effective AI agents in our academy: academy.dair.ai 💬 2 🔄 3 ❤️ 17 👀 1815 📊 7 ⚡