Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy
Correct memories can still push an agent toward agreeing with a user when the context calls for doubt.
The paper proposes MemAdapter to decide how much retrieved long-term memory should shape an LLM agent’s answer. Its three steps test memory risks through counterfactual reasoning, reflect on each memory’s relevance to the current task, and ground the final response in evidence. The authors report gains on three benchmarks for memory reliability across different scenarios. HF Daily Papers' note
The paper proposes MemAdapter to decide how much retrieved long-term memory should shape an LLM agent’s answer. Its three steps test memory risks through counterfactual reasoning, reflect on each memory’s relevance to the current task, and ground the final response in evidence. The authors report gains on three benchmarks for memory reliability across different scenarios. HF Daily Papers' note
score 4