Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware Packing

On September 7, 2026, arXiv published the paper “Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware Packing”, proposing an online clustering and memory pipeline called RSM-full. This approach aims to address the issue of tight prompt word budgets in long-distance large model deployment, combining cosine-gated maximum member merging writing rules with atom-aware grouping context packers. In the AMA-Bench benchmark test, RSM-full achieved 83% full context quality while consuming only 32% of token costs (with a budget of 4k), outperforming the recent streaming clustering baseline Online K-Means by an average of 3.5 to 6.0pp. Ablation experiments confirmed that the performance improvement mainly stems from the merging rules and grouping packers. Additionally, this approach is also effective in the independent benchmark RealMem…