Paper review · November 9, 2026
MedGraph: keeping track of patient state with graph memory in long-term medical conversations
MedGraph: Maintaining Patient State with Graph Memory for Long-Term Medical Dialogue
The authors built a graph memory that chains dated, subject-tagged clinical facts, plus a patient profile, and achieved about 10 percentage points higher accuracy than the strongest baseline on MedMemoryBench.
Problem
An LLM medical assistant that helps a patient across multiple visits must accurately remember the patient's changing state. The authors analyzed 150 wrong answers from five existing memory systems and found three causes of failure. First, the systems rely on outdated state (State validity). Second, they mix in other people's health information (Subject attribution). Third, they miss safety information such as allergies (Safety prioritization).
Method
MedGraph is a three-layer memory made up of raw conversations (L1), a clinical fact graph (L2) and a patient profile (L3). In the memory-building stage, the LLM is called once per session to extract dated facts, and each fact is linked to the person it refers to. Records of medications, diagnoses, tests and vital signs are chained by (item, subject, type), and supersede edges preserve past states. In the retrieval stage, only the patient's facts are ranked with BM25 and dense retrieval, and each chain is treated as a single unit. The record valid at the time the question refers to, the record just before it and excerpts from the raw conversation are included together. The patient profile, which holds all safety facts, is attached to every question without any relevance check.
Results
- In the GPT setting on the 967 questions of MedMemoryBench, it reached an average accuracy of 51.9%. That is 10.5 percentage points higher than the strongest baseline, Embedding RAG (41.4%).
- In the Qwen3-30B-A3B setting, it also averaged 46.1%, 10.4 percentage points higher than Embedding RAG (35.7%).
- On temporal location questions (TLA) it scored 68.4%, well ahead of MemOS (50.3%), and on inference generation questions (IG) it scored 37.9%, well ahead of Embedding RAG (22.2%) (GPT setting).
- The share of evidence mixed with other people's information was 0.0%, lower than the baselines' 12.2–27.3%. The rate of missing required safety facts was also 1.1%, lower than the baselines' 38.2–80.9%. Removing the patient profile raised the miss rate to 49.4%.
- It had the highest average accuracy at all 10 checkpoints. When the same early questions were asked again after the history had grown longer, the performance drop was not significant.
Limitations
The evaluation relies on a single benchmark, MedMemoryBench, and an LLM judge (claude-sonnet-4-5). On multi-step clinical reasoning (MCD), it scored 15.1% in the GPT setting, lower than Embedding RAG (17.2%). On state updates (SUA), it was also lower than Mem0 (63.0%). In the Qwen setting, the BM25-only variant (46.2%) was about the same as the full method (46.1%), so the benefit of adding dense retrieval is unclear. The failure-type analysis is based on a sample of 150 wrong answers.
Why it matters
In fields like medicine, where state keeps changing and missing safety information is dangerous, memory based on simple similarity search is not enough. This paper builds domain requirements such as 'whose fact is it,' 'when is it valid' and 'always surface safety information' directly into the memory structure. It is a useful reference for designing long-term conversational agents.
Key terms
- Long-term Memory
- An external storage structure in which an LLM agent stores conversation content across multiple sessions and later retrieves the evidence needed for a question.
- Supersede chain
- A structure that links records of the same clinical item for the same patient in time order. Even when a new record replaces an earlier one, past states remain, so the state at a specific point in time can be looked up.
- Subject Attribution
- The task of determining whether health information mentioned in a conversation belongs to the patient or to someone else, such as a family member, and linking it accordingly.
- RRF(Reciprocal Rank Fusion)
- A method that combines rankings from different retrieval methods, such as BM25 and dense retrieval, into a single ranking using reciprocal-rank scores.
Original
Authors: Donghua Zhang, Zizheng Wang, Ze Dong, Yiding Feng, Siyu Qian · Posted 2026-11-05 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.