MemOPD: On-Policy Distillation through Memory State Alignment for Long-Horizon Agents
The episode discusses MemOPD, a method for training long-horizon agents with compact memory. Hosts explain that flattening interaction transcripts misaligns training states, causing invalid PPO and distillation supervision. MemOPD reconstructs exact per-call states, uses packing for efficiency, and verifies with RCE. Benchmarks show large F1 gains, especially on longer horizons.