CAMP-II trains smaller agents to save task state in ordinary files, search it and read it after their active context is replaced. A frozen larger teacher scores the smaller models’ own generated actions, including the contents of their notes. This is post-training a policy, not attaching a memory plug-in to an existing agent.
- On five task settings, the trained 4-billion-parameter model averaged 64.6% success. Its untrained counterpart scored 45.3%; a file-memory baseline trained on task rewards scored 54.2%. The tests use 128 tasks per setting.
- When the authors replaced file operations with dedicated memory tools or compaction, average success fell to 48.2%. Masking the teacher’s feedback on memory-writing tokens reduced it to 47.3%; randomly masking the same amount of feedback left 63.3%.
- On network-free test subsets, the trained 9-billion-parameter model scored 38.0% on Terminal-Bench 2.1 and 30.1% on SWE-Bench Pro V2. The untrained 9-billion-parameter model scored 23.1% on each.
The main training setup requires a continuation-file write before a context reset. That requirement cannot explain the whole result: with the requirement removed, the trained policy still scored 62.6%, or 64.0% after retraining without it. Direct memory-swapping tests also show that saved content affects later outcomes. These are within-task findings for two student sizes and one frozen teacher; larger deployed models remain untested.