camel

本文提出DiscernMem，一个免训练框架，通过缺陷感知记忆精炼与边界学习元记忆机制，帮助基于LLM的多智能体系统修正有缺陷的记忆并学习何时正确使用记忆，在AutoGen、CAMEL和MacNet上取得最佳平均分。

The author built a lightweight, drop-in security gate that implements DeepMind's CaMeL principle of preventing untrusted data from authoring actions, achieving ~70% auto-inference accuracy on a benchmark and zero silent unsafe misclassifications, but notes gaps in provenance tracking and robustness.