Building an Enterprise AI Memory System from Scratch
Developer 驻波 presents an introductory architecture for building an enterprise AI memory system. The proposed flow covers asynchronous ingestion, entity linking and topic normalization, BM25 and embedding retrieval, permission filtering, deletion-state checks, versioning, and evidence validation. A Hangzhou travel-guide example illustrates how a prior conversation could be recalled using structured fields. The example and system behavior are not independently validated, and the report does not establish that a complete production system was deployed. The author also says future articles will cover Runtime, evaluation, and vLLM.
The article offers a concrete design for searchable conversation memory while distinguishing source content from durable user facts and preferences. It also shows how normalized entities, canonical topics, permissions, versioning, and evidence checks can constrain retrieval.
Bg
AI已经进入程序员很多日常中。过去一年，驻波从前端转到后端再转向 AI Infra，后面有空写文章时候将分享一些真实的企业级AI开发相关的知识(主播不买课)，Runtime，评测,vLLM,这些都会聊到的。进入正题我们看向本集重点：“Memory系统初步构建”
ShowCase
User:“帮我出一个杭州旅游攻略。”
Output: "这是一份详细的杭州旅游攻略......(Conetxt1)"
---------- 一段时间之后 ---------
User：“上次我们聊天杭州游玩攻略再给我一下。”
Output: "我还记得呢，上次我们说的杭州旅游攻略......(Context1)"
为什么长对话的情况下AI貌似真有记忆，而且比人清楚的多，真实的一字不差的复述以前很久的内容。
行为拆解
第一次输入
这里我们先不分析Runtime对用户Query的操作，当作全部的Query不做加工直接给Memory服务，那么第一次对话。Memory系统接收到的是：