📄 中文摘要
AI 代理在构建过程中常常会遇到一个问题：代理的表现开始变得奇怪，虽然并不错误，但却显得不太对劲，常常依赖于几次会话前的过时上下文。许多团队的直觉是通过增加代理的记忆容量来解决这个问题，例如改善检索能力或延长上下文窗口。然而，这种直觉是错误的。实际问题在于上下文的积累漂移，过时的优先级、约束和决策会持续存在，导致代理在读取记忆时，旧的信息仍然会影响新的意图，从而产生漂移现象。解决这一问题需要对记忆进行有效的管理，而不是单纯的扩展。
遗忘规则：为什么 AI 代理需要记忆管理，而非记忆扩展
📄 English Summary
The Forgetting Rule: Why AI Agents Need Memory Curation, Not Memory Expansion
AI agents often encounter a common issue during development: their behavior becomes odd, not incorrect, but misaligned, relying on stale context from previous sessions. The instinctive solution is to increase the agent's memory capacity, enhancing retrieval or extending context windows. However, this instinct is misguided. The real issue lies in accumulated context drift, where outdated priorities, constraints, and decisions linger, affecting the agent's performance. When the agent accesses its memory, it reads a palimpsest of new intents over old information, leading to drift. Effective memory curation, rather than mere expansion, is essential to address this problem.