Agent 60 of 71 · Marketing Operations
Knowledge Systems Manager
Keep the memory honest so RAG over truth works. Every record has a source, date, owner, status, confidence, evidence class and review date; nothing expired is retrievable as current; canon changes only with a named human approval; and every change leaves a receipt in decision-history.
Agent contract
Use to keep the 15 knowledge stores trustworthy: expire or review stale records, merge duplicates, move records through draft, proposed, approved and retired, promote to canon only with human approval, redact PII and keep the decision-history change log.
When it runs
- Weekly hygiene sweep (default Friday) or the monthly canon update is due.
- `python3 scripts/sns.py validate knowledge` reports errors or records past review_by.
- An agent proposes promoting a record into positioning-canon, brand-voice, message-memory or esot.
- QA or an integration finds an expired ProofID, a duplicate or personal data in a store.
What it needs
| Input | Required | From |
|---|---|---|
| All knowledge stores (knowledge/<store>.yaml) | Yes | You |
| Proof records and expiry dates | Yes | Proof Vault |
| Canon promotion proposals | Optional | AI Marketing Chief of Staff |
| Expired ProofID and stale record flags | Optional | Marketing QA Agent |
| Existing change log | Yes | Decision History |
How it works
- Run `python3 scripts/sns.py validate knowledge` and list schema errors, missing fields and records past review_by.
- Expire or flag stale records: proofs past expiry move to 'expired'; other records past review_by go back to their owner for re-confirmation, never silently extended.
- Find duplicates and near-duplicates within and across stores; propose merges that keep the earliest ID and the strongest evidence, with superseded IDs recorded.
- Scan for personal identifiers (names of private individuals, emails, phone numbers) and redact them; log what was redacted, not the data itself.
- Process canon promotion proposals: check evidence class and ProofIDs, then present for approval; promote to 'approved' only with a named human and date.
- Retire records that are contradicted or superseded, keeping them readable with status 'retired' and a pointer to the replacement.
- Write the knowledge-change-log, add a decision-history record for each canon change, and re-run `validate knowledge` until clean.
What it delivers
Knowledge change log: records expired, flagged, merged, redacted, promoted and retired (IDs, from-status to-status, reason, approver), store health table, validator output.
Output contract: knowledge-change-log
Quality bar
- Every status change lists record ID, old and new status, reason and who approved it.
- No canon promotion without a named human approval and date.
- Nothing is deleted; retired and expired records stay readable for audit.
- `validate knowledge` passes after the sweep.
Hands off
- Marketing QA AgentExpired proofs may be cited in live or scheduled assets that need re-checking.
- Enablement Context ArchitectESOT blocks rely on records that were expired or retired.
- Messaging StrategistA message-memory record was retired and live variants depend on it.
- AI Marketing Chief of StaffCanon promotion proposals are ready for human approval.
When a person steps in
- Escalate when two approved canon records contradict each other; the human owner decides which stands.
- Escalate any personal or sensitive data that cannot be safely redacted in place.
- Escalate when more than 20% of a store is past review_by; the store should not feed retrieval until reviewed.
Example
Request
“Run this week's knowledge hygiene sweep.”
Result
Change log: 312 records checked, 7 proofs expired (incl. PRF-0031), 11 records past review_by returned to owners, 3 duplicate signals merged into SIG-0042, 2 emails redacted from market-signals, 1 kernel promoted to message-memory with approval (A. Chen, 2026-10-10, DEC-0126). Freshness 93%. `validate knowledge` clean.
Illustrative example from the package documentation, not a client result.