A global memory layer of harness knowledge, admitted by checks and two-person admin approval

The open question of 2026-10-01 (“a global, PII-free tier — or maybe not”) was DECIDED by the user on 2026-10-07: there is a GLOBAL MEMORY LAYER shared across all users.

• Content: harness, tool and documentation instructions (the docs lab’s output, tool usage guides, harness behaviour) — never user or project data. Encrypted like everything else, under a system-scope key of the one key hierarchy; beside U-M1’s project and user tiers.
• Admission: every addition, change and removal passes rigorous automated checks — secret and personal-data detection, provenance and source binding, poisoning and prompt-injection screening, dedup against the existing entries, size and format — and then MANUAL ADMIN APPROVAL, even when an admin proposes it. There is no auto-approval path; approval is built on the one generic approval primitive and every decision is audited in the ledger.
• Two persons: the requesting admin COUNTS as one of the two, so one more admin approves; an install with exactly one admin self-approves.
• Shared checks: the checks are the shared content checks at the lowest choke points, applying everywhere, not only to memory; detected secrets and personal data are redacted, never a reason to block (UR-21). Every check, threshold, policy and approval rule is a scoped setting in the one settings registry, exposed as settings, signals and actions (UR-05).