When two of your agent's memories conflict, what decides which one wins?
retained text
Genuine question for people running agents with persistent memory across sessions. Sooner or later two stored facts disagree. The user said they prefer morning meetings, then later said afternoons work better. A runbook says the approval limit is 5k, a newer note says 10k. The retrieval brings back both. Something has to decide. What actually decides in your setup? Is it just recency, last write wins? Source authority, where some entries outrank others? Do you keep both and let the model sort it out at answer time, or does something resolve it at write time? And the part I'm really curious about: do you keep the loser? If the old entry is deleted, you lose the trail of what changed and when. If it's kept, you need a way to mark it superseded so it doesn't keep getting retrieved as if it were current. Asking because conflict handling seems like the part of memory everyone hits and nobody talks about. Retrieval gets all the attention, but the moment two memories disagree, retrieval has already done its job and you're in a different problem.
document fetched 2026-10-24 10:22 · canonical https://reddit.com/r/AI_Agents/comments/1xdkkcw/when_two_of_your_agents_memories_conflict_what