Experiment Approach For Knowledge Graph RAG (Dense-Mem) #5343

I built an open-source MCP memory server called Dense-Mem:

https://github.com/markhuangai/dense-mem

The idea I am trying to explore is that RAG is very useful for retrieval, but durable AI memory may need more than vector search alone.

In daily LLM workflows, the hard problems are often not just "find a similar chunk." They are questions like:

• What did the user actually say?
• Is this evidence, a proposed claim, or an accepted fact?
• Is there a newer fact that supersedes the old one?
• Are two memories in conflict?
• What source supports this answer?
• Can the same memory be reused across different AI tools?

So Dense-Mem separates the host LLM from the memory layer:

• the host LLM handles conversation and judgment
• Dense-Mem stores evidence, typed claims, accepted facts, provenance, conflicts, embeddings, and graph recall
• MCP clients can recall the same durable memory instead of rebuilding context from scratch every session

I wrote more about the motivation here:

https://markhuang.ai/blog/ai-memory-beyond-rag

And I made a hosted demo so people can try it without self-hosting:

https://markhuang.ai/blog/dense-mem-hosted-demo-test-instance

I do not want to overclaim that this "solves memory." I am mostly curious whether this architecture feels useful, overbuilt, or missing something important.

If anyone has thoughts, I would especially appreciate critique on:

1. Does the memory-server / host-LLM boundary make sense?
2. Is graph-backed memory a useful abstraction here, or should this remain simpler?
3. What evaluation would best show whether this improves accuracy in real daily LLM workflows?
4. What failure modes should I test before claiming this direction is useful?

Thanks for reading. I would genuinely value technical pushback.