From text to knowledge object
A loose sentence such as “quotes are valid for thirty days” sounds clear, but it raises questions immediately. Does this hold for every service? Since when? Is it in the terms, a price list or somebody's meeting note? Who is allowed to change it? A memory you can rely on keeps that metadata together with the text.
The minimum fields
- source: file, system, owner and a reference you can find;
- scope: organisation, team, role, customer or process;
- version: content hash, reason for the change and previous version;
- validity: effective date, expiry date and review moment;
- status: proposal, in review, active, conflicting or superseded;
- approval: authorised actor, timestamp and the exact version.
Conflicts are product data
When the CRM, the price list and the contract name different amounts, Jarvis should not average that conflict away. The right output is a review task with three sources, not a confident answer. Conflict status and resolution become part of the audit trail.
Retrieval is authorisation too
Relevance on its own is not enough. A source can fit perfectly in meaning and still not be allowed for that agent, that person or that task. So the server filters first on tenant, role, classification and purpose, and only then ranks on relevance.
Automatic maintenance
A scheduled Codex session can spot expired sources, check links and propose fixes for inconsistencies. It may not replace pricing policy, contract text or customer knowledge on its own. The review task is the automatic result; publication stays human.
