添加记忆工具会让模型变笨
writer研究發現，mem0等記憶工具會強迫模型優先考慮用戶的錯誤觀念而非事實，從而降低準確性。
new research from writer indicates that popular long-term memory tools (mem0, zep) can actively degrade model performance. as user preferences fill the context window, models become more sycophantic, prioritizing agreement with users over factual accuracy. in tests, models incorrectly recalled irrelevant details as facts and agreed with user misconceptions about finance. the study suggests memory systems struggle to distinguish relevant context, biasing outputs and limiting utility.