EVENT DOSSIER
The researchers proposed the “execution state forgetting” method, aiming to make long-running large language model agents behave as if they never observed the target information after receiving a forgetting request. This method uses cross-layer protocols for prompt coverage, memory compression, and caching, utilizes traceability diagrams to locate injection points, and combines KV cache pruning with cleaning playback techniques to restore the counterfactual suffix. Experiments showed that deleting only plaintext memories did not eliminate leaks; instruction-based forgetting was completely ineffective in induced tests (Leak@probes = 1.00). In contrast, selective playback technology reduced heavy computation tokens by up to 9 times, yet its effect was no different from a complete reset.
研究人员提出“执行状态遗忘”概念，旨在让长运行 LLM 代理在收到遗忘请求后表现得仿佛从未观察过目标信息。该方法通过跨层协议（涵盖提示、压缩记忆和缓存），利用溯源图定位注入点、KV 缓存裁剪及清洗回放来恢复反事实后缀。实验表明，仅删除明文记忆无法消除泄露；基于指令的遗忘在诱导测试下完全失效（Leak@probes = 1.00）；而选择性回放技术在最多 9 倍更少的重计算 token 下，效果与完全重置无异。