PERMA is a benchmark for evaluating personalized memory agents in long-horizon conversations where user preferences evolve over time.Instead of static retrieval, models must track event-driven preference evolution and maintain persona consistency under realistic interaction noise.
This dataset supports two complementary evaluation protocols:
