Efficient Memory Management in Multi-Session Dialogue Systems via LLM-Guided Summary Usefulness Prediction

Long-term dialogue systems often suffer from degraded memory quality as accumulated session summaries introduce redundancy, noise, and hallucinated content. Prior work has mainly focused on improving summarization quality, leaving the usefulness of individual summary items largely unexplored. This paper presents a memory management framework that predicts the usefulness of each summary item using pseudo labels generated through an LLM-as-a-Judge procedure and prunes low-value content before it enters the memory. The proposed model follows a two-stage training strategy: it first learns a usefulness prediction module on top of a frozen encoder, then applies LoRA-based adaptation to refine encoder representations for task-specific discrimination. Experiments show that the model more accurately captures the relative importance of summary bullets, and dialogues conditioned on pruned memory exhibit more coherent and contextually appropriate responses. These findings highlight that managing the quality of stored memory, rather than expanding its volume, is crucial for improving long-term dialogue performance.