A compact-memory agent hits 83% of full-context quality at 32% of the token cost, and the gain comes from the merge rule not the retriever

arXiv 2609.04915·medium signalXBlueskyLinkedInCopy link

RSM-full pairs a cosine-gated max-member merge write rule with an atom-aware grouped context packer, and on AMA-Bench at a 4k budget reaches 83% of full-context quality for 32% of the tokens, beating the closest streaming-clustered baseline by 3.5-6.0 points (p<.001) across the roughly 2.6k-5k regime. Ablations attribute +5.7 points to the merge rule over Online K-Means and matched-τ DP-means, and +5.0 points to the grouped packer over flat concatenation. It reproduces on the independent RealMem persona benchmark, beating Budget-RAG, Streaming-Proto and the A-MEM 2025 agentic-memory baseline, though the authors are explicit that higher-token baselines still win outside the 2k-5k window.