MemoryWalker

"Stop Training Agents on Contexts They Never Saw." Alibaba's Token Foundry (Pengjun Xie, Jieping Ye) observes that harnesses which compress context mid-episode, as Claude Code and Qwen-Agent do, mean the model never saw the flat trajectory it is usually trained on; the learning object is a tree. The paper introduces LogitTree, a packed 4D attention mask, and SDCC, a single-backward relaxation, to train on what the agent actually saw.