长时记忆Long-term Memory (Agent Memory)
Long-term Memory (Agent Memory)
让 Agent 跨会话记住用户偏好、任务状态与关键事实,而不是每次从零开始。Let an agent remember user preferences, task state and key facts across sessions instead of starting from zero each time.
通常分三层:工作记忆(当前上下文)、情节记忆(历史事件)、语义记忆(提炼后的事实与偏好)。
写入要有策略:什么值得记、何时提炼、如何过期与删除;否则记忆会变成噪声。
与「上下文窗口」不同:长时记忆是外部存储与检索,不受单次上下文长度限制。
召回也要有门槛:相似度不等于相关,错误记忆比没有记忆更难排查。
It usually has three layers: working memory (current context), episodic memory (past events) and semantic memory (distilled facts and preferences).
Writes need policy: what is worth remembering, when to distil, how to expire and delete; otherwise memory becomes noise.
Unlike a context window, long-term memory is external storage and retrieval, not limited by a single context length.
Recall also needs a threshold: similarity is not relevance, and a wrong memory is harder to debug than no memory.
2026 年的 Agent 都在卖「记得你」;不懂记忆的写入与召回,就无法判断它是真记得还是临时拼接。Agents in 2026 sell remembering you; without understanding how memory is written and recalled, you cannot tell real memory from a temporary stitch-up.
Agent Memory · 智能体记忆
回到名词地图看第3段的其他 75 个词 →Back to the map: the other 75 terms in stage 3 →