← sources

There's a reason why AI memory is still fucked

source: reddit (reddit:AI_Agents)published 2026-09-26 16:11ingested 2026-09-27 02:35

retained text

Every memory tool is just a fkin 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 for production 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?

document fetched 2026-09-27 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wqudww/theres_a_reason_why_ai_memory_is_still_fucked