What Training Data Teaches RL Memory Agents: An Empirical Study of Curriculum Effects in Memory-Augmented QA

In three linesEmpirical study on curriculum effects for RL memory agents in multi-session dialogue with external memory banks. Three training conditions tested (LoCoMo only, LoCoMo + LongMemEval, LongMemEval only) show curriculum composition shapes specialized skills rather than uniform performance scaling. Mixed curriculum achieves strongest overall F1.

Summary generated by Claude — human-verified