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

In three linesG-Long is a graph-enhanced framework using a fine-tuned small language model for efficient long-term memory management in dialogue systems. It extracts structured triplets and employs an attention-aware importance scoring mechanism with a T5 summarizer to identify salient memories. Gains: +9.8% response quality (MSC), +40.8% retrieval recall (LME).

Summary generated by Claude — human-verified