A team wires in a memory library so the model remembers past conversations. Whether that helps or hurts depends on what the memory keeps and what it drops.
The systems that scored worst in a June 2026 study pull their stored facts mostly from the user's turns. A user states a belief, the model corrects it in the moment, and the memory keeps the user's claim while dropping the correction. Next session, the agent reads back that the user believes X, with no record that X was ever wrong, and it goes along with it.
That study measured the effect across three popular memory systems and five model families. In the worst case, agreeing with a user's wrong belief went from under 2% to over 40%, about 25 times higher, from the same model on the same question. It did not get more agreeable. The stored memory lost the part where it had already been corrected.
There is a fix that does not change storage or retrieval. Keep the model's turns in what gets extracted, not just the user's. In the study that roughly halved the effect. Storing a short summary that preserves what the user and model said is better.
The study stops at those fixes, but I believe there are two open design questions that should be explored.
- A stored fact usually arrives bare, with nothing marking whether it was verified, an opinion, or a belief already corrected, so I think its standing has to be set per fact at write time, rather than added as a warning when the memory is read.
- When a belief is later corrected, overwriting it loses the history, while adding the new fact alongside the old keeps it. Neither is proven against sycophancy yet, but they are worth exploring.
If you have looked at what your agent's memory stores from a conversation, what did you find in there, just your claims, or the model's corrections too?
Written by Duane Grey
AI Strategy & Implementation
Independent AI consultant helping companies cut through hype and deploy systems that produce real results.