上海AI Lab团队推出了MemHarness框架，旨在使大语言模型Agent能够像人类一样灵活地调用和重构过往经验 [1]。该框架通过在检索与动作之间插入显式的记忆重构环节，实现了Agent记忆的动态调整 [1]。
MemHarness框架由检索、重构和动作生成三个环节组成 [1]。研究团队采用GRPO进行端到端优化，优化信号来自任务奖励和格式约束，无需人工标注即可进行训练 [1]。在ALFWorld和WebShop基准测试中，该方法的性能显著优于纯强化学习基线和静态记忆增强方法 [1]。采用该框架的7B规模模型表现超过了Gemini-2.5-Pro等闭源模型 [1]。此外，MemHarness在分布外场景中展现出强大的鲁棒性 [1]。
The Shanghai AI Lab team has unveiled MemHarness, a framework that enables large language model agents to dynamically reconstruct and retrieve past experiences in a manner similar to human memory processes [1]. The approach inserts an explicit memory reconstruction stage between information retrieval and action generation, allowing agents to flexibly adapt their learned experiences to new situations [1].
MemHarness operates through three sequential components: retrieval, reconstruction, and action generation [1]. The framework employs GRPO for end-to-end optimization, drawing training signals from both task rewards and format constraints, without requiring manual annotation [1]. In comparative evaluations on the ALFWorld and WebShop benchmarks, MemHarness significantly outperforms both pure reinforcement learning baselines and static memory-augmented methods [1]. Notably, the 7-billion-parameter model variant demonstrates performance exceeding that of closed-source models such as Gemini-2.5-Pro [1].
The framework exhibits robust generalization capabilities in out-of-distribution scenarios, suggesting practical applicability across varied task environments [1].
