Skip to content- Sources: primary
- Summary: The authors argue that prior on-policy distillation for agent memory relied on custom memory tools the teacher never saw in pre-training, so its guidance was weak. CAMP-II instead gives the policy plain Bash and a persistent workspace across five CAMG-II environments and distills file-based memory use from a frozen teacher with fully asynchronous OPD. It reports 2.97 times the gain of the strongest OPD baseline, with improved success on Terminal-Bench 2.1 and SWE-Bench Pro V2.
- Why it matters: Agent builders get evidence that teaching memory through ordinary file operations, which models already know, transfers better than bespoke memory interfaces.
- Background: On-policy distillation samples trajectories from the student and has a teacher score or correct each step, which avoids the distribution mismatch of training only on teacher outputs. Long-horizon agents outgrow their context window and must compact history, so what they write down before a reset decides what they can recover.