What is an LLM memory layer under the hood
I'm struggling to understand what applications like Mem0 actually do. Are they just an api layer that allows you to add information to a vector database?
When I go to the Mem0 website their explanation is pretty convoluted but to me it just seems to be a way to easily integrate rag and nothing else. I was wondering if anyone here can shed some light as to what it really is and if there's more to it.
5 Replies
The way I understand and anyone can correct me if I’m wrong but it’s basically a way to structure and customize the “memory” layer. LLMs are stateless so they aren’t actually learning anything. Tools like mem0 and others just structure the memory that the llm can access selectively. Think of it like context7 mcp but you can choose what’s available for the llm to call
this is the paper i found on their website, its very detailed https://arxiv.org/pdf/2504.19413
this is perfect thank you
So I'm understanding it as mem0 being more of a classic rag while mem0g using a knowledge graph and both use some sort of name-entity recognition or something similar to add relevant information to their databases