Every engagement starts the same way: with a wrong answer somebody can point at. What follows is the work of finding which part of the record produced it, and making that part answerable to someone.
testimonial 1 — a quote, verbatim, with permission
testimonial 2 — a quote, verbatim, with permission
testimonial 3 — a quote, verbatim, with permission
You have agents in production and no way to answer what they are actually reading, or which of it is stale.
I trace what each system retrieves, where those facts were first written down, and which of them contradict each other. The output is a register: every source, its owner, its freshness, and what breaks if it is wrong.
Your agent forgets between sessions, or remembers something you corrected weeks ago.
Deciding where memory lives is a cost decision before it is a technical one — what it costs to write, to retrieve, and above all to correct. I design that layer against your actual sources rather than a reference diagram.
Nobody owns the facts your AI acts on, so nothing supersedes anything and wrong answers never get retired.
Who owns a fact, what replaces it, and how a wrong one is withdrawn — set up as a working practice your team runs without me. This is the part that decides whether the first two survive contact with a real organisation.
If you can describe the wrong answer, that is usually enough to work out which part of the record produced it. No preparation needed.