RSM-full hits 83% LLM quality at 32% of token cost in tests
TL;DR
- RSM-full reaches 83% of full-context quality at 32% of the token cost at a 4k budget on the authors' AMA-Bench.
- On RealMem, it beats A-MEM by +1.65pp and Streaming-Proto by +2.97pp; ablations credit +5.7pp to the merge rule and +5.0pp to the packer.
- The authors say gains hold only in a 2k–5k prompt-token band; higher-token baselines remain stronger outside that range.
Under a 4k token budget, RSM-full reaches 83% of full-context quality at 32% of the token cost on AMA-Bench, the authors' compact-memory benchmark. On an independent test called RealMem, it beats A-MEM by +1.65pp and Streaming-Proto by +2.97pp, according to the arxiv preprint from Jiahe Geng, Jinpeng Wang and Kun Yuan.
The design pairs a "cosine-gated max-member merge write rule" with an "atom-aware grouped context packer". Three-seed ablations attribute +5.7pp of the improvement to the merge rule and +5.0pp to the packer over flat concatenation.
"Across benchmarks the message is consistent: under tight budgets, compact-memory performance is driven mainly by how streaming memories are merged and how retrieved content is assembled," the authors write. They caveat the scope directly: RSM-full "is most useful when answering roughly 2k–5k prompt tokens," and higher-token baselines still win outside that band. On BM25-RAG parity they explicitly decline the equivalence claim, calling the paired delta of +0.27pp not significant at p=.47.
Two experts in our Who's Who directory shared the preprint.
Shared on Bluesky by 2 AI experts
Originally reported by arxiv.org
Read the original article →Original headline: Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware Packing