BMT-05.03 Executive Summary#
BlueMirror.tech | May 2026#
A medication management app continued recommending dosage timing based on a prescription discontinued eight months earlier. The patient followed the outdated recommendation for three weeks before a pharmacist caught the discrepancy. The system had remembered too much. It served the person she was, not the person she is.
BlueMirror treats forgetting as a first-class engineering problem. The architecture implements structured temporal decay where different types of context become stale at different rates, governed by domain-specific rules rather than uniform deletion policies.
Active medications decay very slowly, on the order of years, because clinical relevance persists even after discontinuation. The fact that the person took metformin for three years remains relevant to her drug interaction profile indefinitely. Shopping preferences decay fastest, on the order of weeks, because brand preferences shift with availability and experience. Relationship status never decays automatically, because people matter until the person says otherwise. Cognitive assessment history never decays, because the longitudinal trajectory itself is diagnostic. Each domain carries a half-life: after one half-life, the confidence weight is halved; after two, quartered; below a configurable threshold, the context is flagged or archived.
The system forgets in three distinct ways. Temporal decay is automatic and continuous: context confidence degrades over time based on domain-specific rates. It is reversible; if the person re-demonstrates a decayed preference, the confidence restores immediately. Event-triggered forgetting is semi-automatic: a medication change triggers immediate deactivation of the old prescription’s active status and propagates updates across all concierge agents that reference it. One event, multiple agents updated, coordinated through the shared MoC schema. Person-directed forgetting is manual: the person tells the system to forget something specific, and the system complies immediately, with a confirmation step and a plain-language explanation of what will be lost.
The architecture separates active context from historical record. Active context lives in MoC Layers 0 through 3 and decays over time. Historical record lives in archival storage and retains indefinitely unless the person explicitly requests deletion. This separation means the system can forget for daily performance without losing for clinical safety. The buying agent stops suggesting metformin-era dietary adjustments three months after the medication change. But when the physician asks about medication history, the health concierge retrieves the full record from archival storage.
Person-directed deletion goes further than archival. When the person requests deletion, the data is purged from active context, archival storage, and cached copies across all agents. For data shared externally, the system cannot delete what it does not control, but it notifies the external party, logs the notification, and removes its own copy. The audit trail records the deletion event without preserving the deleted content.
The full article is available at BlueMirror.tech.