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

Version 0.11. This version adds a second, larger controlled measurement and a full account of what an external audit of the implementation changed. It supersedes 10.5281/zenodo.20819013 (v0.10, June 2026), which remains available. When a long-running AI agent fills its context window, the harness summarizes everything said so far and continues from that summary. The summary is good at the gist and bad at the specifics: file paths, version numbers, identifiers, and decisions made moments earlier get smoothed away. This paper is about stopping that, cheaply, without retraining the model. Our central finding is a calibration-density law: the fidelity with which an agent's own facts survive compaction scales with how densely those facts are emitted as explicit, confidence-marked claims. We exploit this with a training-free PreCompact rehydration hook that harvests the agent's own calibration-marked claims into a verbatim brief and re-injects it after compaction, augmenting the native summary rather than replacing it. New in this version. A controlled equal-budget measurement over 69 real compactions and 1,491 probes frozen before scoring, against a control arm the system wrote itself: the brief recovers 33.9% of probes against the native summary's 18.0% (+15.9 points). The result is deliberately reported as a trade rather than a total — bare prose facts rise from 7.3% to 29.4% while file paths fall from 72.2% to 29.9%, because a rendered brief spends characters on structure and fewer items fit the same budget. Also new: an audit section. The implementation measured in versions up to 0.10 did not do what the paper described. An external audit found defects across the harvesters, the budget, the storage layer and the benchmark itself — including a confidence marker being applied to an entire message rather than its own statement, and text authored by other speakers being stored as the agent's claims. All are fixed and described; every figure produced before the fix is dated as such. Correcting the benchmark alone moved the headline from +31.7 to +15.9 points. The contribution is an honest, reproducible benchmark, a density law for continuous memory, a never-hallucinate augmentation layer, and a worked example of an audit correcting a system's account of itself — not a universal >90% breakthrough. Implementation, tests and full defect history: github.com/rushnur88/claimkeep. The benchmark reproduces both tables on any Claude Code transcript directory.