MedExpMem lets medical AI learn from its own diagnostic mistakes, boosting accuracy 7%

Current medical vision-language models (VLMs) rely on static parameters that cannot evolve across patient encounters — they lack the experiential learning that human physicians use to refine differential diagnoses. A team of researchers from multiple institutions (including The Chinese University of Hong Kong) introduces MedExpMem, a novel framework that equips VLMs with an experience memory system. Unlike retrieval-augmented generation (RAG), which only fetches static disease descriptions, MedExpMem actively stores pairwise differential notes derived from the agent's own diagnostic failures. These notes contain key discriminators, actionable decision rules, and reasoning error patterns — effectively mimicking how doctors learn from mistakes.

The framework follows a two-phase learning process: initial practice exposes knowledge gaps, then reflective re-diagnosis refines understanding. On a radiology benchmark covering 11 subspecialties, MedExpMem consistently improved accuracy across diverse model architectures and scales, with gains up to 7.0%. Analytical experiments confirmed memory quality and robustness. The paper, accepted as an early accept at MICCAI 2026, highlights MedExpMem as a competitive method for medical adaptation beyond what parametric learning alone can achieve.