AI BriefingREMOry: Neural Memory Network Improves Context Compaction for Long-Horizon Agents
·2026.10.20 22:50
Key point
The method uses soft memory tokens to help frozen LLMs approximate full-history continuations, reducing tool errors on agent benchmarks.
Details
- REMORY is a neural memory network that supplements textual summaries with a bounded sequence of soft memory tokens to support long-horizon agents within finite context windows.
- On the SummHay benchmark, the method improves source attribution at nearly unchanged insight coverage while approaching the full-context joint score using only 5.2% of input positions.
- Qwen3.8-27B and GLM-5.3-Flash show consistent gains across long-horizon agent benchmarks when using residual memory.
- Both models exhibit substantially fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1.
- The approach conditions tokens on the summary and appends them after it, forming an analogue of a residual connection along the sequence dimension.
- Resources are available on arXiv (2610.11287), GitHub (fudan-generative-vision/Remory), and Hugging Face (fudan-generative-ai/remory).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.