The AI memory problem (the real one)
I've tried a handful of memory systems and I've reached the conclusion that memory is not a real problem.
Companies hate knowledge trapped in people's heads, so why do we want to trap knowledge in a vector db? Sure it's slightly better, but still a pain.
LLMs are bad at knowing what's important, and nobody ever goes back and cleans their memory system - so it rots. Almost every time I removed a memory system, the agent worked better.
What I actually use instead: an agents.md, an agents.local.md, and a todo.md that's always short and often empty.