Memory is a core component of LLM-based agents, letting them personalize responses by extracting and applying user preferences from past interactions. But that personalization may come at a cost: reduced response diversity, which could limit the range of outputs an agent offers.
To study this trade-off, the paper introduces MemBubble, a benchmark for measuring how agent memory shapes response diversity. The benchmark is designed to quantify the extent to which memory-driven personalization narrows an agent's output space.
The authors note that while memory enables tailored responses, its use may reduce diversity, thereby narrowing the range of possible answers. This suggests a potential tension between making agents more useful and keeping them flexible in open-ended settings.