LLM agents ignore correct memory facts, raising reliability concerns

A developer testing a support agent found that even when memory retrieval returned the correct fact (user on enterprise plan), the LLM still suggested upgrading. This failure mode differs from forgetting: the fact was in context but the model used it incorrectly. Popular memory frameworks like Mem0, Zep, and Letta don't verify whether responses reflect retrieved facts.