Mitigating Over-Personalization in LLMs via Structured Memory

2026-08-08

AI Summary

This work addresses the challenge that large language models, when personalized using long-term memory, often suffer from cross-domain information leakage and sycophantic behavior due to unstructured memory injection, compromising response authenticity and appropriateness. To mitigate this issue, the authors propose a lightweight inference-time intervention strategy that structures long-term memory into domain-specific partitions and selectively presents relevant segments instead of injecting the full memory context. This approach requires no modifications to either the model architecture or the stored memory content. Evaluation across seven prominent large language models on the PersistBench benchmark demonstrates that the method reduces cross-domain leakage by 8.8% on average while preserving task-relevant utility, thereby significantly alleviating over-personalization without sacrificing performance.

Abstract

Conversational assistants increasingly rely on persistent long-term memory to personalize responses across sessions. However, when stored user information is reintroduced into the model context, it can also influence responses in inappropriate or unrelated settings. We study two such failure modes in memory-augmented LLMs: cross-domain leakage, where memories from one life domain affect responses in another, and memory-induced sycophancy, where stored user beliefs make models more likely to agree with the user rather than respond truthfully. We apply a simple inference-time modification to how memories are presented to the model, without changing the model or the memory contents. Across seven models on PersistBench, we compare the commonly used all-in context format, where memories are injected as an unstructured list, with structured formats that partition memories by domain. This simple modification consistently reduces cross-domain leakage while preserving utility, with our strongest method reducing leakage by $8.8\%$ on average relative to the baseline.

Problem

• over-personalization
• cross-domain leakage
• memory-induced sycophancy
• structured memory
• LLMs

Innovation

• structured memory
• cross-domain leakage
• memory-induced sycophancy
• personalization
• inference-time modification