Brief, mostly facts
REMORY learns residual memory to compact agent context
An arXiv paper adds soft memory tokens to context compaction summaries. It uses 5.2% of input positions and cuts repeated tool calls and errors on two agent benchmarks.
What the research found
REMORY (arXiv 2610.11287) adds soft memory tokens to context compaction summaries, reaching near full-context performance on SummHay with 5.2% of input positions and fewer repeated tool calls and errors on two agent benchmarks; code and weights are linked, but it is a research paper.
Claims and how well the evidence supports them
- REMORY approaches full-context joint score using 5.2% of input positions on SummHay 98%
- REMORY reduces repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1 97%
- Paper is arXiv 2610.11287 with code on GitHub 98%
Evidence
- arxiv.org https://arxiv.org/abs/2610.11287
- github.com https://github.com/fudan-generative-vision/Remory
Actionability rated medium.