REMORY adds soft residual memory tokens to context compaction, hitting near full-context scores at 5.2% of input
Hanchen Xia · hf · 2026-10-09
- REMORY is a neural memory network for long-horizon agents: after compacting history into a textual summary, it appends a bounded sequence of soft memory tokens that help a frozen LLM approximate what it would produce with the full history—analogous to a residual connection along the sequence dimension.
- On SummHay, it improves source attribution at nearly unchanged insight coverage and approaches 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, with substantially fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1.
More from coding & agent
- Codex turns vague install instructions into real charges once a card is stored — CurieuxExplorer · 2026-10-09
- LLM Wikis: Let an AI Curator Build Growing Knowledge Bases in Your Obsidian Vault — dSebastien · 2026-10-09
- readdown-mcp turns web pages into token-efficient, LLM-optimized Markdown — modelcontextprotocol · 2026-10-09
- MCP-native marketplace lets agents discover and pay for APIs with USDC credits — modelcontextprotocol · 2026-10-09
- OTP expiry exposes browser agent's weakest link: human-in-the-loop auth — sujingshen · 2026-10-09
- PiChat: An Autonomous Agent Living in Your iPhone with Durable, Resumable Execution — aigclink · 2026-10-09