Create a multi-agent diagnostic framework that keeps a persistent memory of 'reasoning history' across patient cases to improve future accuracy without full retraining.