Last week in London, we had a memorable evening at the 𝗨𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗗𝗮𝘁𝗮 ...
TL;DR · AI 摘要
Milvus团队在伦敦非结构化数据聚会分享AI Agent构建经验，强调上下文管理、记忆系统与向量检索基础设施的关键作用。
核心要点
- 单纯提升大模型能力不足以构建实用Agent，需融合企业文档、用户偏好等多源上下文
- Agent记忆应是可编辑、可检索、自更新的动态系统，而非静态日志
- 生产级Agent依赖高并发、低延迟、支持元数据过滤的向量检索基础设施
结构提纲
按章节快速跳转。
思维导图
用一张图看清主题之间的关系。
查看大纲文本（无障碍 / 无 JS 友好）
- AI Agent基础设施需求
- 上下文管理
- 企业文档
- 用户偏好
- 历史对话
- 记忆系统
- 可编辑文件
- 摘要生成
- 向量化检索
- 自动清理
- 检索基础设施
- 低延迟
- 多租户
- 元数据过滤
- 成本优化
金句 / Highlights
值得收藏与分享的关键句。
Better models are not enough. Agents need the right context, from enterprise documents and user preferences to past conversations and external knowledge.
Agent memory is a living retrieval system. It turns raw conversations into editable files, summaries, embeddings, and searchable memory, then keeps that memory clean as context changes.
Once agents start working, retrieval infrastructure becomes the bottleneck. Production systems need to handle scale, latency, multi-tenancy, metadata filtering, and cost-efficient vector search.
A big thank you to everyone who joined us. The thoughtful questions, technical discussions, and hallway conversations made" / X
Last week in London, we had a memorable evening at the 𝗨𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗗𝗮𝘁𝗮 𝗠𝗲𝗲𝘁𝘂𝗽 𝗟𝗼𝗻𝗱𝗼𝗻, alongside Ying Hou and Firas Jarboui. A big thank you to everyone who joined us. The thoughtful questions, technical discussions, and hallway conversations made the evening even more valuable, and brought real-world depth to the conversation. During the meetup, our Head of Developer Relations at Zilliz, Jiang Chen, shared his experience building better agents, and why stronger models alone are not enough. Key takeaways from the talk: Better models are not enough. Agents need the right context, from enterprise documents and user preferences to past conversations and external knowledge. 
Agent memory is a living retrieval system. It turns raw conversations into editable files, summaries, embeddings, and searchable memory, then keeps that memory clean as context changes. 
Once agents start working, retrieval infrastructure becomes the bottleneck. Production systems need to handle scale, latency, multi-tenancy, metadata filtering, and cost-efficient vector search, which is where Milvus and Zilliz Cloud become part of the agent stack. 
Watch the full replay: youtube.com/watch?v=3mDFw9#Milvus#VectorSearch#AIAgents