Cross-session memory in AI companions is a much harder problem than I expected

I've been experimenting with persistent memory for AI chatbots and wanted to share something that surprised me about how hard the retrieval problem actually is everyone talks about memory for chatbots like its solved - throw things in a vector db, do cosine similarity, inject into context. but theres a massive gap between "storing facts" and "surfacing them at the right moment in conversation" the storage part is easy. the hard part is knowing WHEN to retrieve a memory example: in one session a user casually mentioned trying to cook risotto and burning it.