MIT 与 Meta 论文 ReHarness：用后续任务结果修订智能体记忆笔记
MIT 与 Meta 等机构的论文提出 ReHarness，在每个任务结束后由修订智能体依据笔记原始任务的事实收窄或修正被检索出的旧笔记，解决过时经验持续误导智能体的问题。在 3 个基准和 3 个前沿模型上平均归一化增益 0.240，最佳原始轨迹基线为 0.168；在 CL-Bench 与 Claude 上重放完整原始历史要多消耗 3.4 倍 token。
Rohan Paul· @rohanpaul_ai · X·2026-11-13 08:39· 53 分钟前AI 评分50
AI 导读
MIT 与 Meta 等机构的论文提出 ReHarness，在每个任务结束后由修订智能体根据该笔记原始任务的事实收窄或修正被检索出的旧笔记，解决笔记一旦写成就很少更新、过时经验持续误导智能体的问题。在 3 个基准和 3 个前沿模型上平均归一化增益 0.240，最佳原始轨迹基线为 0.168；在 CL-Bench 与 Claude 上重放完整原始历史要多消耗 3.4 倍 token。
正文 · 原文
New MIT and Meta paper finds that agent memory works better when old notes get corrected using what happened the next time they were used.
Agents that learn on the job save notes from past tasks and reuse them. Once written, those notes rarely change, so an overly broad or outdated lesson keeps misleading the agent.
After each task, ReHarness sends the retrieved notes and the result to a revision agent. It narrows or corrects each note, using only facts from that note's original task.
Across 3 benchmarks and 3 frontier models, it averaged 0.240 in normalized gain, the share of possible improvement over a memoryless agent. The best raw-trajectory baseline scored 0.168, and replaying full raw history used 3.4× as many tokens on CL-Bench with Claude.
If your agent learns from experience, let each outcome edit the notes that shaped it, not only add new notes.
来源：Rohan Paul · x.com
来源：RadarAI · aihot.news