Continuous Memory for Multi-Agent Infrastructure: A Calibration-Density Law for Surviving Context Compaction

ClaimKeep is a training-free PreCompact augmentation hook for long-horizon multi-agent LLM systems. Before context compaction it harvests the agent's own confidence-marked claims and feeds them back, augmenting native compaction without replacing it (never-worse by construction). Across two independent scorers and two corpora it produced zero confident-wrong answers, with a union product-lift of +8.3 strict and +16.6 semantic over native compaction. The central finding is a calibration-density law: continuous-memory fidelity scales with the density of explicit confidence-marked claims. Code, tests, and a reproducible benchmark: https://github.com/rushnur88/claimkeep (MIT).