MedGraph: Maintaining Patient State with Graph Memory for Long-Term Medical Dialogue
MedGraph beat the strongest tested baseline by 10.5 accuracy points with a GPT backbone and 10.4 with Qwen on MedMemoryBench.
The authors identify three recurring memory failures: outdated patient states, health facts attributed to the wrong person, and missing safety information. MedGraph dates facts, assigns them to the person they describe, and links clinical records into chains that preserve earlier states. Retrieval reads those chains at the time asked about and includes the original dialogue text. A patient profile supplies every recorded safety fact with each question.
ArXiv · AI/CL/LG's note
The authors identify three recurring memory failures: outdated patient states, health facts attributed to the wrong person, and missing safety information. MedGraph dates facts, assigns them to the person they describe, and links clinical records into chains that preserve earlier states. Retrieval reads those chains at the time asked about and includes the original dialogue text. A patient profile supplies every recorded safety fact with each question.
ArXiv · AI/CL/LG's note
score 4