More memory, worse agents: why saved lessons can make AI agents repeat mistakes

A new preprint finds evidence that agents storing their own rules can lock in errors. Keeping raw logs and writing fresh guidance per task scored higher and cost fewer calls.