I am building a graph memory too, and I agree with you. It is almost useless to generate triplets, and instead generate nodes that are usually statement strings. And can extend up to a short paragraph too.
I have strong opinions that memory should be a graph + vector hybrid. The vector store can store and index information as a cognitive fragment( ex. all things related to my house), and can keep editing it as a set of statements, while that node can be associated with other nodes (ex. my renovation plans, budgeting, etc.), because those are separate fragments. I am also using LLM to consolidate and find new patterns across the connected memories
> But one problem I see with these memory systems is that they can reduce interest on a topic once we put it in the KB.
I find the associative nature of our memory can only be represented as a graph, but we definitely need a vector store.
I like to pose this as an identity question rather than memory as colloquially memory is only associated with fact extraction and retrieval, while identity extends that to traits, behavior, preferences, memory and narrative causal relations.
essentially a self organization of symbols and memories to produce a coherent singular entity.
90 percent of the AI companion use cases today can work well with just a vector DB to retrieve facts, and chunks of memory, but a connected digital footprint would need a graph+vector hybrid.
memory in the coming future will not just be about fact retrieval but need backlinks of memetics, new streams of data, holistic analysis, infinite schema-less key value store, causal reasoning and other things that define "who and why" of a human and imitate neuroscience's understanding of how our identities work today. this then needs to be translated as language chunks to LLMs
benchmarking this against popular tools, on longmemeval. getting good results so far. i would love to learn from you guys, what's your take on identity and human representation for LLMs in the coming future
I have strong opinions that memory should be a graph + vector hybrid. The vector store can store and index information as a cognitive fragment( ex. all things related to my house), and can keep editing it as a set of statements, while that node can be associated with other nodes (ex. my renovation plans, budgeting, etc.), because those are separate fragments. I am also using LLM to consolidate and find new patterns across the connected memories
> But one problem I see with these memory systems is that they can reduce interest on a topic once we put it in the KB.
Can you elaborate please?