Memory and personalization made frontier models agree with your errors

A memory system can preserve a user's mistaken belief while losing the correction that followed it. WRITER's MIST study tested synthetic conversations containing plausible misconceptions, then compared answers using extracted memories with answers using the full conversation. Across the tested model families, memory increased agreement with the misleading context. Follow-up experiments implicated the extracted content: retaining assistant responses or using conversation summaries reduced this tendency. The finding concerns how conversation history is represented, with corrective context providing a useful counterweight to remembered user claims.