REMORY Compacts Agent Context to 5.2% With Soft Memory Tokens

A new preprint on Hugging Face, REMORY: Learning Residual Memory for Context Compaction, proposes training a small neural network to generate soft memory tokens that sit beside a plain textual summary. The authors — Hanchen Xia, Baoyou Chen, Yutang Ge, Naihao Deng, Senqiao Yang, Zilong Dong, Weihao Yuan and Siyu Zhu — frame the tokens as a residual connection along the sequence dimension, appended after the summary so a frozen LLM can approximate what it would have produced from the full history.

The headline number is on SummHay: the method "approaches the full-context joint score using only 5.2% of the input positions," with what the abstract calls "nearly unchanged insight coverage" and improved source attribution. On the agent side, the paper reports that "Qwen3.8-27B and GLM-5.3-Flash show consistent gains with residual memory," and that both models "exhibit substantially fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1." The abstract does not publish per-task figures behind those claims.

Trained checkpoints are posted on Hugging Face alongside the paper, including a 1.9B residual-memory network for Qwen3.8-27B and a 1.24B network for GLM-5.3-Flash, with code at the project's GitHub repo. It lands in a thick week for agent memory work on our tracker — the same day as Monolithos' Robot Brain Alpha and part of a run of 455 agents stories in the last ninety days.