代理记忆仍然是一个未解决的问题

2作者: manthangupta10911 天前原帖
如今,许多初创公司声称他们已经“解决了代理记忆”。 在许多情况下,代理记忆被简化为一个带有嵌入和检索的向量数据库。这并不是代理记忆,而仅仅是存储加搜索。 真正的代理记忆涉及以下方面: - 代理应该记住什么与忘记什么 - 记忆如何随时间变化(衰退、强化、巩固) - 何时应隐式与显式地读取记忆 - 记忆如何影响规划、工具选择和行为 - 如何处理冲突、过时或依赖上下文的记忆 向量数据库可以是系统的一部分,但并不是整个系统。 在代理能够可靠地推理自己的记忆,而不仅仅是检索信息块之前,我们仍处于早期阶段。 我很好奇其他人的看法: - “解决代理记忆”究竟意味着什么? - 这是否只是一个单一的问题,还是一系列未解决的问题的集合?
查看原文
A lot of startups today claim they have &quot;solved agentic memory.&quot;<p>In many cases, agentic memory is being reduced to a vector database with embeddings and retrieval. That’s not agentic memory. That’s just storage + search.<p>Real agentic memory is about: - What an agent should remember vs forget - How memory changes over time (decay, reinforcement, consolidation) - When memory should be read implicitly vs explicitly - How memory influences planning, tool choice, and behavior - How to handle conflicting, stale, or context-dependent memories<p>A vector DB can be part of the system, but it’s not the system.<p>Until agents can reliably reason about their own memory and not just retrieve chunks. We’re still in early innings.<p>Curious what others think: - What does &quot;solving agentic memory&quot; actually mean? - Is this even a single problem, or a bundle of unsolved ones?