能不能提前给他灌输一些记忆，然后当客服用

[sh308178685]

RT

[freshpomelo]

我感觉可以把准备好的记忆以历史对话的形式传进来，然后evermemos处理成自己的格式，这个就要消耗LLM token了。
然后如果要当客服，还得再套一个chatbot吧，不能直接用。EverMemOS的出发点是LLM的基础设施，暂时不能直接2C

[gaogzs]

现在来看EMOS只擅长读入对话形式的记忆，而且储存时是以第三方视角分别对各个用户进行评估。我也希望能支持一种特殊的记忆形式，专门用来储存特定agent需要的基础知识

[freshpomelo]

现在来看EMOS只擅长读入对话形式的记忆，而且储存时是以第三方视角分别对各个用户进行评估。我也希望能支持一种特殊的记忆形式，专门用来储存特定agent需要的基础知识

我没太明白你说的基础知识的定义范围，“基础知识”是与用户无关，但与agent的应用场景相关？
如果是我说的这种情况，那是不是应该在你agent的system prompt里？

[gaogzs]

现在来看EMOS只擅长读入对话形式的记忆，而且储存时是以第三方视角分别对各个用户进行评估。我也希望能支持一种特殊的记忆形式，专门用来储存特定agent需要的基础知识

我没太明白你说的基础知识的定义范围，“基础知识”是与用户无关，但与agent的应用场景相关？ 如果是我说的这种情况，那是不是应该在你agent的system prompt里？

对，是这个意思。但假如agent需要的知识量特别大（比方说几千条往上），一次性塞system prompt会浪费输入token，也有很大可能想不起来。我想利用EMOS每次根据用户的输入提取最相关的那几条知识，动态地加进prompt里

[freshpomelo]

你这个场景的确得想想，现在好像不太支持

[gaogzs]

你这个场景的确得想想，现在好像不太支持

我现在是在试着把这些知识以独白对话和模拟对话的形式用常规手段传入记忆。但目前感觉下来就是memcell切割地很粗糙，很难提取到真正相关的记忆

[freshpomelo]

“memcell切割地很粗糙”能给一些具体的例子吗。
因为现在EMOS是会对原始对话进行总结概括只保留少部分必要的细节。
我理解你的场景是要保留大量基础知识的细节，然后查询时返回最相关的那部分的细节，和现在EMOS的处理逻辑其实并不能很好的兼容。

[rookiedali]

搞客服 为啥不直接用rag

[freshpomelo]

搞客服 为啥不直接用rag

这是个好问题。客服其实有两个刚需：1. RAG静态知识库。2.客服面对的用户的聊天动态上下文。EMOS以记忆的形式提前处理好了聊天动态上下文，比直接给LLM更省token，也更快，也支持更久的对话历史。
如果实现了agent级别的静态知识库共享，那就可以不用RAG了。而且RAG有个比较大的问题是多跳查询效果很差，EMOS在多跳上通过agentic_retrieve效果会好很多

[ZhenhangTung]

I think the discussion here has been very valuable. My view is that we should identify the respective strengths of Memory and the Knowledge Base, and combine them in a way that ultimately improves retrieval accuracy, which is the metric everyone cares about most.

Regarding the thousands of data entries mentioned by @gaogzs, my suggestion is to store them in the Knowledge Base rather than in Memory, and expose them as Tools for the Agent to call—whether for RAG or Agentic RAG.

So what role does Memory play in the Agentic RAG pipeline? I think this is something worth discussing openly.

In customer support scenarios, for example, you can preload certain customer information and business context (which could also exist as part of the system prompt). My current thinking is that Memory can influence the reasoning process of Agentic RAG, helping the agent use the Knowledge Base tools more effectively.

That said, these are still abstract ideas. The concrete implementation really depends on community developers co-building around specific use cases.

[cyfyifanchen]

Closing as part of the EverOS 1.0 issue triage. The repository has been refactored around the 1.0 Markdown + SQLite + LanceDB architecture and current use-case/docs structure. PR #258 documents the migration boundary: #258. If there is a specific 1.0 code or docs change needed, please open a fresh focused issue.

[cyfyifanchen]

Reopening after a stricter post-triage audit. This is a general use-case question, not a purely pre-1.0 bug report. EverOS 1.0 may support this through explicit memory ingestion, but the customer-service/pre-seeding workflow still deserves a proper answer or doc pointer instead of being closed as obsolete.