G-Long: Graph-Enhanced Memory Management for Efficient Long-Term Dialogue Agents

G-Long is a proposed framework that enhances memory management for long-term dialogue agents by employing a fine-tuned small Language Model (sLM) for structured triplet extraction and associative retrieval, addressing the inefficiencies of existing methods. It introduces an attention-aware importance scoring mechanism that utilizes cross-attention signals from a T5 summarizer to prioritize salient memories. Experimental results indicate G-Long achieves state-of-the-art performance, improving response quality by up to 9.8% on the MSC benchmark and retrieval recall by 40.8% on LME, while significantly reducing computational costs, making it a valuable tool for practitioners focusing on efficient long-context reasoning in dialogue systems.