TBNAI
REMORY: Learning Residual Memory for Context Compaction
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 07/10/2026, 20:00:00 UTC
Briefing created: 15/11/2026, 13:53:00 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 09:18:24 UTC.
Briefings
REMORY is a neural memory network that adds soft memory tokens to LLM summaries to improve long-horizon agent decision-making. It reduces input requirements by 5.2% while maintaining near-source attribution and reducing tool errors on benchmarks like BrowseComp and Terminal-Bench 2.1.
Key points
- REMORY adds soft memory tokens to LLM summaries to form a residual connection along the sequence dimension.
- Qwen3.8-27B and GLM-5.3-Flash models show consistent gains with REMORY on long-horizon agent benchmarks.
Context
Long-horizon agents compact history within finite context windows, but textual summaries alone may not support all subsequent decisions. REMORY supplements summaries with bounded soft memory tokens to improve decision quality.
Limitations
Evidence covers only SummHay benchmark; results on BrowseComp and Terminal-Bench 2.1 are reported but specific error reduction metrics are not detailed.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 15/11/2026, 13:53:00 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiffmpzeg6mnc4q3vsto2qygpfohwoe56jo7wvq7sx5c6nemo56zs4