Last week our assistant cited a decision we made three years ago. Correctly, with a link to the exact note it came from. That moment took two weeks of pipeline work and three years of accumulated notes — and the pipeline is the part worth writing down.
This is the system we run on our own six businesses. Nothing in this post is a concept diagram.
Where the knowledge actually was
The same place yours probably is: spread across tools and heads. Notes in different apps, decisions buried in email threads, runbooks that existed only as habits. Running six businesses made it worse, not better — every project re-solved the same context problem, and every AI session started from zero.
One governed vault
Step one was unglamorous: consolidate everything into plain markdown files with YAML front-matter, linked into one graph, versioned in git. No proprietary format, no SDK, no translation layer — files that any editor and any model can read.
The vault syncs to every device through a self-hosted, end-to-end encrypted channel. The sync server only ever holds ciphertext, so on that layer we could not read the notes in transit even if we wanted to. Capture a thought on a phone and it is in the graph everywhere in seconds.
The manifest: BigQuery and Knowledge Catalog
Here is the part people conflate, so let’s be precise. The vault publishes a manifest — a structured listing of every note and its metadata — into BigQuery. Knowledge Catalog registers that manifest automatically, so every note becomes a governed, discoverable asset: what exists, where it lives, how it is tagged, who owns it.
That layer does governance and discovery, full stop. It does not ground anything, and it does not serve content to any model. When an agent tool needs to know whether a runbook exists, the catalog answers. When a human asks a question, a different layer entirely does the work.
The grounding layer: a search data store
Grounding happens in a separate, dedicated search data store attached to Gemini Enterprise. The full vault indexes into that store, and Gemini Enterprise grounds its answers on the index. Ask “what did we decide about pricing in March?” and the answer comes from our own history, with citations that link to the exact notes it drew from.
Two layers, two jobs. The catalog is for governance; the data store is for grounding. Keeping them separate keeps each one auditable — and removable. Indexing for the grounded assistant is opt-in per tier, into a dedicated, scoped data store that can be deleted without touching the vault itself.
1,038 documents, zero failures
Our last full sync indexed 1,038 documents with zero import failures. Zero is not luck; it is a property of the format. Plain markdown with YAML front-matter either parses or it does not, and validating it is cheap enough to run before every sync. There is no export from a proprietary tool to go subtly wrong, no rich-text edge case to swallow half a document. Boring formats fail loudly and rarely — which is exactly what you want underneath an index your answers depend on.
Answers that write back
A cited answer is useful once. A cited answer that saves back into the vault is useful forever. Answers can write back as new notes, linked into the graph, so the next question grounds on the last one’s conclusion. The graph compounds instead of resetting.
Scoped views that fail closed
Not everyone should see everything. Each person or group gets a gated web assistant on its own subdomain, grounded only on the slice of the vault they are allowed to see. The scope is enforced server-side and fails closed: a query outside it does not return a redacted answer — it simply finds no sources. We run a scoped external-partner view in production today, and the partner cannot retrieve anything outside their slice.
Agents on the same brain
The same store serves our agents. The morning brief that lands at the start of each day cites the graph. Agents run policy checks against the vault before they act — the standing rules live as notes, so “check policy first” is a retrieval step, not a hope. Human or agent, nobody works from memory anymore.
See it running
The gated assistant is live — see the assistant page for how scoped access works. Or book a free 30-minute walkthrough of the whole pipeline: the graph, the sync, the cited answers, on the real system rather than a demo environment.
