Memory layer for AI agents is totally FUCKED
retained text
I'm building an agent and needed a memory layer so I tested 4 memory SDKs before picking one and ran the same test on 4 memory tools, told the agent I prefer dark mode in session 1 then in session 5 told it I switched to light mode then in session 8 asked it which mode I use. Every one gave me both answers back and the model just picked whichever one matched the query better so sometimes it said dark mode and sometimes light mode depending on how I asked. None of them got rid of the old fact when the new one came in. One just added the new one next to the old one, another saved both separately with nothing connecting them, one gave me the newer one most of the time but asking differently brought the old one back and the last one needed me to go in and change it myself because it doesn't deal with contradictions at all. I tested entity resolution too, mentioned my name vansh in one session and vansh from india in another and VS in a third. One made three separate entries, one connected the first two but missed the shorthand, one thought all three were different people and one just doesn't do entity resolution at all. These are the two most basic things memory should handle and none of them got either one right. If you're looking at any memory tool please give it a fact that changes across sessions and mention the same person three different ways and if it can't figure out either then it's just searching through old conversations and returning whatever looks closest. also please help me find something that handles both.
document fetched 2026-09-26 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wpzjri/memory_layer_for_ai_agents_is_totally_fucked