Structured Memory Reduces Over-Personalization in LLMs

A recent preprint on arXiv (2608.08300) examines how enduring long-term memory in conversational agents can result in excessive personalization, leading to unsuitable replies in unrelated situations. The study highlights two main issues: cross-domain leakage, where memories from one area of life influence responses in another, and memory-induced sycophancy, where retained user opinions prompt models to agree with users instead of providing honest answers. The authors recommend a straightforward adjustment during inference: rather than presenting memories as a disorganized list, they suggest organizing them by domain into structured formats. This approach, which does not alter the model or the memory data, shows a consistent reduction in both cross-domain leakage and memory-induced sycophancy across seven models tested on PersistBench. The findings indicate that the way memories are formatted for the model significantly affects its responses, providing an effective strategy to lessen over-personalization in memory-enhanced LLMs.