REMORY: Learning Residual Memory for Context Compaction
The paper proposes soft “residual” memory tokens to preserve details that summaries lose during context compaction.
REMORY adds a learned neural memory after a textual summary, letting a frozen LLM approximate what it would have done with the full history. The authors report better source attribution on SummHay while using 5.2% of the input positions. In long-horizon agent benchmarks, Qwen3.8-27B and GLM-5.3-Flash both improved with the residual memory setup. The paper also says the models produced fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1. HF Daily Papers' note
REMORY adds a learned neural memory after a textual summary, letting a frozen LLM approximate what it would have done with the full history. The authors report better source attribution on SummHay while using 5.2% of the input positions. In long-horizon agent benchmarks, Qwen3.8-27B and GLM-5.3-Flash both improved with the residual memory setup. The paper also says the models produced fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1. HF Daily Papers' note
score 5