MedExpMem: Adapting Experience Memory for Differential Diagnosis
A paper proposes MedExpMem, a memory framework that lets medical vision-language agents learn from their own diagnostic mistakes. It reports accuracy gains of up to 7.0% on a radiology benchmark.
Intelligence analysis by GPT-5.4 Mini
MedExpMem is designed to give medical AI something closer to physician experience, not just stored facts. Instead of retrieving generic disease descriptions, it saves discriminative lessons from failed diagnoses and uses them to improve later differential reasoning.
A medical AI can be good at naming things, but still get confused when two problems look almost the same. This paper says the AI should also keep notes about the mistakes it made, like a student learning from wrong answers.
Those notes are not big encyclopedia pages. They are more like little clues, such as "this one looks similar, but the key sign is different." That is like remembering how to tell twins apart by one small feature.
The paper says this helped the AI do better on a radiology test. The biggest improvement was 7.0%, which means the memory trick made the system better at choosing between confusing medical cases.
Analysis
What the paper proposes
MedExpMem is an experience memory framework for VLM-based diagnostic agents. The paper argues that experienced physicians do more than recall disease knowledge: they also learn how to separate confusing conditions over time. The authors say current medical VLMs do not evolve across encounters in that way, because their parameters stay static.
Rather than using retrieval-augmented generation to pull in broad disease descriptions, MedExpMem stores discriminative experience from the model's own diagnostic failures. The memory is organized as pairwise differential notes that capture key discriminators, actionable decision rules, and reasoning error patterns. In other words, the system tries to remember what went wrong and what helped distinguish one case from another.
How it is built and tested
The framework uses a two-phase process that mirrors physician learning. First comes initial practice, where the agent reveals gaps in its knowledge. Then comes reflective re-diagnosis, where the agent revisits the case and refines its understanding. When a new case appears, the agent retrieves relevant experience memory to support differential reasoning.
The paper evaluates MedExpMem on a radiology benchmark covering 11 subspecialties. According to the abstract, the method delivers consistent accuracy gains across different models and scales, with a maximum improvement of 7.0%.
Takeaway
The main claim is that this kind of memory is useful for medical adaptation in a way that parameter updates alone do not provide. The authors present MedExpMem as a competitive approach for improving differential diagnosis by turning past mistakes into reusable guidance.
Key points
- MedExpMem stores experience from the model's own diagnostic failures instead of only static medical knowledge.
- The memory is structured as pairwise differential notes with discriminators, decision rules, and error patterns.
- The method uses a two-phase process: initial practice followed by reflective re-diagnosis.
- On a radiology benchmark spanning 11 subspecialties, the paper reports accuracy gains of up to 7.0%.
