RSM-full paper hits 83% of full-context quality at 32% token cost for memory-limited LLM agents
rohanpaul_ai · x · 2026-09-11
An arXiv paper (2609.04915) introduces RSM-full, an online clustered-memory pipeline for long-horizon LLM agents facing tight prompt budgets. Key idea: organize the past before optimizing retrieval — group related memories and keep them together rather than retrieving isolated chunks.
- Two designs: a cosine-gated max-member merge write rule plus an atom-aware grouped context packer.
- Results: on AMA-Bench it reaches 83% of full-context quality at 32% of token cost under a 4k budget, beating Online K-Means by +3.5–6.0pp (p<.001) across the 2.6k–5k regime. Ablations attribute most gains to the merge rule (+5.7pp) and grouped packer (+5.0pp).
- Replication: improvements reproduce on the independent RealMem benchmark, significantly beating Streaming-Proto (+2.97pp) and Budget-RAG (+0.69pp), roughly on par with BM25-RAG.
Authors: Jiahe Geng, Jinpeng Wang, Kun Yuan.
Related event: RSM Achieves 83% Memory Quality with 32% of Tokens(3 posts)→
More from coding & agent
- CTF eval design under fire: prompts turn agent evals into a bizarre meta-eval — voooooogel · 2026-09-11
- DeepsecBench security leaderboard: GPT-6 Astra tops at 37.79, Opus 5 costs $128 per run — JohnPhamous · 2026-09-11
- One Dollar Audit Offers AI Smart Contract Security Audits for $1 on Base — seanwbren · 2026-09-11
- SWE-Together Update: Claude Fable 5 Tops Coding Benchmark, Muse Spark 1.3 Is 5x Cheaper — shuchaobi · 2026-09-11
- GPT-6 Astra 3D Workflow: Blender MCP for Hard-Surface, TripoAI for Organic Models — majidmanzarpour · 2026-09-11
- Gemini Canvas Turns Any Google Sheet Into an App, Sparking Startup-Killing Concerns — VishnuNath · 2026-09-11