REMORY: learned memory tokens help agents keep details after their history is summarised

When AI agents run long tasks, they compress their history into a summary to stay within the model's memory limit, and details get lost. REMORY, a paper featured on Hugging Face Daily Papers, adds a small set of learned "memory tokens" after the summary that help an unchanged model act as if it still had the full history. The authors report that on the SummHay test it improved source attribution with nearly unchanged coverage, approaching full-history results using only 5.2% of the input. With Qwen3.8-27B and GLM-5.3-Flash on long agent benchmarks, they report consistent gains and fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1. The memory tokens are not human-readable, the authors confirmed on the paper page, so you cannot inspect what was kept. This is research, not a product.