Last week we watched our AI cite a decision we made three years ago. Correctly. With the source.
That moment took two weeks to build and three years to earn. This post is about the gap between those two numbers — and why most companies are stuck on the wrong side of it.
AI amnesia
Every AI tool you use starts from zero. No memory of your decisions, your clients, your hard-won lessons. The models are extraordinary. The context is missing.
You have probably said one of these three things yourself.
“I asked ChatGPT the same question I asked it three months ago.” Every session starts from a blank page. Your research, your decisions, your context — gone the moment the chat ends. So you re-explain, re-paste, re-decide.
“When she left, fifteen years of process knowledge walked out with her.” A client said that to us last month. Nobody disagrees with it. Everybody has a version of it. The business runs on knowledge that lives in two people’s heads, and onboarding takes months because nothing is written down where anyone can find it.
“We’re paying for Copilot, Gemini and Claude — and none of them know our business.” Your agents can’t see your data schemas, your runbooks, your naming conventions. Every AI project re-solves the same context problem from scratch.
The models aren’t the problem. The missing piece is a governed place where your knowledge lives — one your AI can actually ground on.
The fix: one vault, grounded, cited
Here is the shape of the system we run.
First, the vault. Everything we know — decisions, runbooks, pricing history, lessons that cost real money — lives in one linked graph of plain markdown files. Human-readable in any editor. Owned by us, forever. Self-hosted, end-to-end encrypted sync puts it on every device; the sync server only ever sees ciphertext.
Then, the grounding. The vault’s manifest lands in BigQuery, where Knowledge Catalog registers every note — that layer is governance and discovery, so each note is a catalogued, queryable asset. Separately, the notes index into a search data store that Gemini Enterprise grounds on. That second layer is what changes the day-to-day: ask “what did we decide about pricing in March?” and the assistant answers from our history, citing the exact notes it drew from. No citation, no claim.
Answers save back into the vault, so the graph compounds. And scoped views mean a partner sees only their slice — enforced server-side, fail closed. A query outside the scope simply finds no sources.
The result is an AI that doesn’t guess what we think. It quotes us back to ourselves.
We built it because we needed it
This isn’t a concept deck. It’s the production system behind our own six-business portfolio — 1,038 documents indexed for Gemini Enterprise grounding, with zero import failures on the last full sync, and 100% of answers grounded and cited back to the exact note.
Six businesses, one graph. The electrician brand’s quoting rules, the pool company’s pricing policy, the agency’s delivery playbooks — all in the same governed vault, all answerable in plain language, all with receipts.
And because the files are plain markdown in a git repository, there is no lock-in to regret later. The whole vault exports with a single git clone. Grounded-assistant indexing is a choice, not a default — it’s opt-in per tier, into a dedicated, scoped, removable data store.
Your AI doesn’t need to be smarter
It needs to know you.
That is the whole thesis. The teams getting compounding value from AI this year aren’t the ones with the cleverest prompts. They’re the ones whose knowledge is structured, governed, and sitting where their AI can ground on it.
If you want to see what that looks like running for real, we do a free 30-minute walkthrough of the live system — the graph, the sync, the assistant that knows three years of decisions. Book a walkthrough, or start with the film at vault.emergedigital.com.
