Number 4 of 12 · November 13, 2026

Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation

Teaching SI agents to use plain files as memory, through a teacher that already knows file tools, gave 2.97 times the gain of the strongest baseline on the authors' new test suite.

What they found

The authors study agents that work on tasks so long that the history no longer fits in the model's active context. These agents must decide what to write down, what to read back, and when. Earlier work trained this with on-policy distillation (OPD, where a student learns from a teacher on the student's own attempts), but used special memory tools the models had never seen in pre-training. The authors argue this makes the teacher's advice weak and gives the student few useful memory actions to learn from. They build CAMG-II, a set of long-horizon environments (Shop, Coding, AutoResearch, General, Visual) that offer ordinary Bash access and a workspace that persists across the episode, so files act as memory. Their method, CAMP-II, trains one policy across all five environments with fully asynchronous OPD from a frozen teacher that already knows file operations. On CAMG-II it gets 2.97 times the gain of the strongest OPD baseline, and it improves success rates on Terminal-Bench 2.1 and SWE-Bench Pro V2.

Why it matters

Long tasks are a main limit for SI agents, because they lose track of what happened earlier. This work suggests that using familiar tools like files and Bash may work better than inventing new memory interfaces. The abstract gives no absolute scores, model sizes or details of the baselines, so how large the improvement is in practice is unclear.

How it changes the race

If the result holds, builders of coding and research agents could get better long-horizon memory from tools models already know, instead of designing custom memory systems. This looks like a practical, incremental step in training methods rather than a new capability.

The abstract, by the authors

Language-model agents increasingly tackle long-horizon tasks whose interaction histories exceed the model's active context. They must therefore decide what to write down and what to read back, and when to do either. Recent work has begun to use on-policy distillation (OPD) to make memory control part of the policy, relying on memory tools and compaction protocols that lie outside either model's pre-training. However, this weakens the distillation on both sides: the teacher cannot give useful guidance on an interface it has never used, and the student samples too few meaningful memory actions for that guidance to take effect. To study this problem, we introduce Coding Agent Memory Gym II (CAMG-II), a suite of long-horizon agentic-RL environments spanning Shop, Coding, AutoResearch, General, and Visual. Alongside each environment's native task interface, CAMG-II provides ordinary Bash access and an episode-persistent workspace, enabling the policy to create, revise, search, and reuse files as memory across context replacements. We also introduce Coding Agent Memory Post-Training II (CAMP-II), which trains a single policy jointly across all five environments with fully asynchronous OPD to learn this file-based memory behavior directly from a frozen teacher with a prior for file operations. On CAMG-II, CAMP-II obtains 2.97 times the gain of the strongest OPD baseline. The memory behavior CAMP-II learns generalizes to two external benchmarks, Terminal-Bench 2.1 and SWE-Bench Pro V2, with improved success rates.

Quoted from arXiv:2611.14009. The research and its words belong to its authors.

The authors

Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong, Yudong Guo, Xinrui Wang, Jiayin Zhu, Gu Simiu, Sulong Xu, Cong Fang

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.