Why is AI memory still all messed up ??

​ Every memory tool is just a fucking vector store with a wrapper on top, you embed the conversation and put it into a database and when you need something it pulls back whatever looks most similar. I build agents on Claude's API and this is what happens with users, one of my users told the agent they moved from delhi to mumbai in session 4 and in session 7 the agent recommends a restaurant in delhi because the old embedding scored higher. User changes a preference and the store keeps both versions and the model picks between them randomly and none of this is a model problem it's the memory not knowing what changed. I need something that does typed extraction so it knows a fact from a preference from an event, also handles contradiction so old facts get replaced and entity resolution so one person mentioned 10 ways stays one entry. what are my options?