MedGraph: Maintaining Patient State with Graph Memory for Long-Term Medical Dialogue
MedGraph organizes longitudinal dialogue as dated, person-attributed chains and retrieves them at the time a question asks about. The paper reports higher accuracy than five baselines on a medical dialogue benchmark.
MedGraph stores dated, person-attributed medical facts and links medications, diagnoses, and measurements into chains. Retrieval ranks patient chains, selects the state relevant to the question's time, and adds original dialogue text; a patient profile supplies recorded safety facts. For builders of agents with persistent memory, the design highlights three failure cases to guard against: outdated information, facts about the wrong person, and missing safety details. The paper reports **10.5 and 10.4 points** above the strongest baseline on MedMemoryBench with GPT and Qwen backbones, respectively. These results are specific to longitudinal medical dialogue. The material does not establish performance in other domains or provide details about deployment with real patient data.
No separate implication has been recorded yet.
MedGraph stores dated, person-attributed medical facts and links medications, diagnoses, and measurements into chains. Retrieval ranks patient chains, selects the state relevant to the question's time, and adds original dialogue text; a patient profile supplies recorded safety facts. For builders of agents with persistent memory, the design highlights three failure cases to guard against: outdated information, facts about the wrong person, and missing safety details. The paper reports **10.5 and 10.4 points** above the strongest baseline on MedMemoryBench with GPT and Qwen backbones, respectively. These results are specific to longitudinal medical dialogue. The material does not establish performance in other domains or provide details about deployment with real patient data.
MedGraph adds a domain-tested design for persistent memory that preserves dates, people, relationships, and relevant dialogue when reconstructing state. It sharpens general source-linked memory patterns with explicit safeguards against stale, misattributed, or incomplete facts, though its reported gains apply to longitudinal medical dialogue.