跳到主要内容Skip to main content

Agent Memory(智能体记忆)Agent Memory

Agent Memory

  • 智能体Agents
  • 第三段 · 用深Stage 3 · Going Deeper
  • 引用核验 2026-11-21refs checked 2026-11-21
解决什么问题What problem it solves

让 agent 在跨会话、跨任务时保留有用的经历、事实和做法,并在需要时检索、更新、遗忘,而不是每次从零开始。Let an agent retain useful experiences, facts and procedures across sessions and tasks, then retrieve, update and forget them instead of starting from scratch each time.

这是什么What it is

Agent Memory 通常分工作记忆、情景记忆、语义记忆与程序性记忆:分别对应当前上下文、过去交互、结构化事实和可复用技能。

完整链路不止「存」,还包括写入筛选、索引与检索、冲突处理、巩固与遗忘;只加一个向量库不等于有记忆。

Agent memory is usually split into working, episodic, semantic and procedural memory: the current context, past interactions, structured facts and reusable skills.

The full pipeline is more than “store”: it includes write filtering, indexing and retrieval, conflict handling, consolidation and forgetting. Adding a vector database alone is not memory.

为什么需要知道Why it matters

长任务和多轮协作里,重复问同一件事、忘记上次的约束、把过期信息当事实,都会直接降低可用性。记忆是 agent 从「一次性问答」走向「持续协作」的关键一层。In long tasks and repeated collaboration, asking the same question again, forgetting a previous constraint or treating stale information as fact quickly destroys usability. Memory is the layer that moves an agent from one-shot answering to sustained work.

别名Chinese aliases

智能体记忆 · 长期记忆 · Agent Memory

出处Sources
  • Memory for Autonomous LLM Agents · arXiv 2603.07670Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers ↗
  • A Survey of Agent Memory · arXiv 2602.06052A Survey of Agent Memory in the Second Half ↗