CueMem addresses the challenge of utilizing extended dialogue histories for long-term conversational agents. Directly using the full history is computationally expensive and prone to inaccuracies. The research proposes a framework that treats extracted memory records as retrieval cues, reconstructing the necessary context from their source turns. During memory construction, CueMem extracts fine-grained cues and links them to their originating turns. At query time, the system retrieves relevant cues, maps them to source turn anchors, and expands from these anchors using a turn graph that considers temporal proximity and semantic relatedness. This process reconstructs a compact evidence context for LLM answer generation.
Experiments on the LoCoMo and LongMemEval datasets show consistent outperformance against established long-term memory baselines. Specifically, CueMem reduces query-time input tokens and latency compared to the full-history LLM setting. The turn graph-based reconstruction method effectively recovers supporting dialogue evidence. The system’s design prioritizes efficient context retrieval over storing the entire dialogue history.
Further analysis indicates that retrieval cues provide a viable alternative to self-contained memory evidence. The framework’s design focuses on reconstructing context rather than storing it, leading to improved performance and reduced computational costs. This approach is particularly relevant for deploying conversational agents in resource-constrained environments.
Source: https://arxiv.org/abs/2609.12354
