New memory pipeline RSM-full preserves context at 32% token cost

A pre-print on arXiv introduces RSM-full, a memory pipeline built for LLM agents that must operate long-term under tight prompt budgets. The problem is familiar: as conversations grow, latency, cost and context-window limits make a full prompt impractical. The question is not just how much to remember, but which memory design delivers the best quality-per-token ratio in the compressed regime. RSM-full combines two mechanisms — a cosine-conditional peer-max merging write rule and an atom-aware context packer in clustering — and reports a strong Pareto point in the 2,000–5,000 token range.

Behind the numbers

On AMA-Bench, the authors' primary benchmark for compressed-memory regimes, RSM-full reaches 83% of full-prompt quality at 32% of the token cost at a 4k budget. Against the nearest streaming baseline, Online K-Means, it leads by 3.5–6.0 percentage points (p < .001) across the entire 2.6k–5k span. A three-seed ablation shows most of the gain comes from the merging rule (+5.7 pp over Online K-Means and DP-means with matched τ) and the cluster packer (+5.0 pp over flat concatenation). Results are averaged over four seeds, adding statistical stability.

Validation on an independent benchmark

The pattern repeats on RealMem, an independent long-horizon persona-memory benchmark. There RSM-full improves Budget-RAG by 0.69 points (p = .006), sits in statistical parity with BM25-RAG (pairwise difference +0.27, p = .47 — the authors stress they do not claim formal equivalence), and beats Streaming-Proto by 2.97 points and the reproduced 2025 agentic baseline A-MEM by 1.65 points (p < .001). The recurring conclusion: under tight budgets, performance is determined mainly by how memories are merged in-stream and how retrieved content is assembled.

Limitations and broader context

The paper states explicitly that RSM-full is effective mainly in the 2k–5k token range; outside that regime, higher-budget baselines remain stronger. It is a pre-print that has not yet undergone peer review, and code, data and demos are available alongside the paper for reproduction. The distinction between "open weights" and "open source" applies here — availability enables inspection, but not necessarily a fully permissive license. For developers building long-horizon agents with hard cost and latency constraints, the work offers a concrete direction: invest in the merging rule and context packing before expanding the budget.