Hugging Face Releases REMORY: Revolutionizing Context Compaction for Long-Horizon Agents
Summary:Hugging Face has introduced REMORY, a neural memory network designed to address the challenge of history compaction for long-horizon agents within a finite context window. REMORY supplements textual summaries with a sequence of soft memory tokens, enabling a frozen LLM to better approximate the continuation it would produce with the full history. In the SummHay benchmark, REMORY significantly improved source attribution while maintaining nearly unchanged insight coverage, approaching the full-co
Background and Challenge
Long-horizon agents often face the challenge of compacting their history within a finite context window to continue processing tasks. However, relying solely on textual summaries may not support every subsequent decision, creating a bottleneck for agent performance.
REMORY's Innovation
Hugging Face's REMORY addresses this challenge through the following innovations:
- Soft Memory Token Generation: REMORY generates a sequence of soft memory tokens conditioned on the history and summary, appending them after the summary to form an analogue of a residual connection along the sequence dimension.
- Approximate Inference with Frozen LLM: REMORY enables a frozen LLM to better approximate the continuation it would produce with the full history, enhancing the agent's decision-making capabilities.
- Efficient Resource Utilization: In the SummHay benchmark, REMORY approaches the full-context joint score using only 5.2% of the input positions, demonstrating significant efficiency gains.
Experimental Results and Performance
When integrated with models like Qwen3.8-27B and GLM-5.3-Flash, REMORY demonstrates the following advantages across various long-horizon agent benchmarks:
- Improved Source Attribution: In the SummHay benchmark, REMORY significantly improves source attribution while maintaining nearly unchanged insight coverage.
- Reduced Repeated Tool Outputs and Errors: In benchmarks such as BrowseComp and Terminal-Bench 2.1, REMORY reduces repeated tool outputs and tool errors.
Industry Impact and Future Outlook
The release of REMORY provides a new technical pathway for the development of long-horizon agents, particularly in scenarios requiring efficient history compaction and precise inference. Its innovative soft memory token mechanism is expected to drive improvements in AI agent performance in complex tasks.
Developer Recommendations
- Integration with Existing Models: Developers can integrate REMORY with existing long-horizon agent models to enhance their history compaction and inference capabilities.
- Stay Updated: Hugging Face may continue to optimize REMORY, so developers should stay updated with the latest releases to benefit from new features.
— END —Source: Hugging Face Daily Papers (2026-10-08)
Tags: #Hugging Face #REMORY #Long-Horizon Agents #Context Compaction #LLMs & Foundation Models
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments