Title: Adapting Memory Retrieval to Evolving User Preferences

Abstract:

In the realm of long-context dialogue systems, determining both the timing of memory access and the relevance of specific interaction history segments remains a critical challenge. Current methodologies often depend on heuristic retrieval cues or maintain continuous memory availability, yet they frequently overlook the dynamic and sometimes contradictory nature of user preferences. To address this limitation, we introduce a cohesive framework for memory access and selection that explicitly accounts for shifting preferences.

Rather than depending on surface-level semantic similarity, we treat personalized memory retrieval as the process of pinpointing historical dialogue turns that offer evidence regarding a user’s underlying preference state. We evaluate the utility of each memory turn through a Bayes factor, which measures the increase in the model’s likelihood for the reference response when a particular turn is integrated into the context. This approach yields a rigorous metric for evidence strength and serves as a singular signal for both accessing and selecting memory.

By conceptualizing memory retrieval as a utility estimation problem, the model acquires the ability to detect significant turns and modulate its memory usage according to expected utility. Our experiments across four diverse memory benchmarks demonstrate that this method surpasses conventional embedding-based retrieval techniques in long-context, preference-heavy tasks where capturing evolving preferences is crucial. Furthermore, the approach maintains strong performance in low-density scenarios where semantic similarity remains adequate.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC