Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy

Researchers introduced Memadapter, a counterfactual adaptation method designed to prevent memory-induced sycophancy in LLM-based agents that retain information across tasks and sessions. The work addresses how persistent memory can cause agents to over-align with users' historical beliefs even when those beliefs are incorrect.

Long-term memory enables LLM-based agents to retain and reuse information across tasks and sessions, supporting personalization and long-horizon interactions. However, persistent memories can also induce sycophancy, causing agents to over-align with users' historical beliefs even when they are inacc