📄 中文摘要
AI 代理的记忆系统存在严重问题，尤其是在长时间运行后。经过 30 多天的持续运行，发现向量相似性并不等同于事实的准确性。在运行的第二周，代理开始依赖过时的上下文信息，检索返回的高相似度匹配往往是事实错误，导致代理基于过时的信息自信执行任务，直到出现故障才被发现。实际案例中，代理错误地认为客户偏好电子邮件沟通，反映出记忆架构的脆弱性。
我们是如何为 AI 代理记忆构建漂移检测（以及为什么仅依靠嵌入失败）
📄 English Summary
How We Built Drift Detection for AI Agent Memory (And Why Embeddings Alone Fail)
The memory system of AI agents faces significant issues, particularly after prolonged operation. After running for over 30 days, it was discovered that vector similarity does not equate to factual accuracy. In the second week of operation, agents began relying on outdated context, with retrievals yielding high-similarity matches that were often factually incorrect. This led to agents confidently executing tasks based on stale information, only to be noticed when something broke. A real example highlighted that the agent incorrectly stored the fact that the client preferred email communication, showcasing the fragility of the memory architecture.