A knowledge base almost always fails at the same point: it grows faster than it gets reread. Past a few hundred documents, retrieval costs more than rewriting, and the value — endlessly deferred — collapses. AI doesn't fix this by tidying up. It fixes it by making the corpus machine-readable — provided it was structured for that. I write from a base I've maintained since 2025: 8,934 documents, including 1,565 synthesis notes and 211 sourced state-of-the-art reviews, queryable by an assistant that answers from what's written there, not from the web.
Why note-taking apps aren't enough
Proprietary apps optimize for capture, not AI recall. Opaque database, closed format, keyword search: a model can neither browse the whole nor reason over it. You stack up knowledge you never mobilize. And the day the tool shuts down or doubles its price, the corpus is captive — and that data is an asset AI doesn't commoditize.
The framework: three layers, not one tool
The reference model was formalized by Andrej Karpathy in 2026, under the name LLM wiki. It holds in three layers, and it's the separation that matters:
- Raw sources, immutable. Documents, meeting notes, transcripts. The model reads them, never writes to them.
- The compiled wiki. Synthesis notes per entity and per concept, written and maintained by the model, cross-linked. Each new source updates ten to fifteen notes — it's the weaving that creates the value, not the stacking.
- The schema. A short file that tells the model how the base is organized, where to look, what it's allowed to edit.
The format itself holds no surprises: plain text, light front-matter metadata, a stable folder tree. Readable by any model, today and in ten years.
What my own base taught me
I measured mine against this framework in September 2026. The result isn't flattering, and that's exactly why it's useful.
- The foundation is solid: the schema, the routing rules, the assistant's memory. That's where most bases fail; mine holds.
- The wiki is flat. 1,565 notes linked to each other, but each produced in isolation: a new source creates one note, it doesn't update fifteen. The weaving is missing — and it's the main lever.
- Search is keyword-only. No semantic layer. I'd set the threshold at 800 notes before installing one; it's been crossed. The engine chosen, open-source and local, exposes the base to the model without any document leaving it.
- The nightly automation depended on a machine staying on. It stopped on May 1, 2026 without warning. The full story is in The AI that dreamed.
Two illusions, in passing. "Too many documents and it'll choke": false — 8,934 text files weigh a few dozen megabytes. And "AI will tidy up for me": also false. Left without guardrails, it bloats the text base into plausible-sounding padding. The rule that holds: the model proposes, the human validates, everything is versioned, and only one process writes at a time.
A knowledge base is only worth what you can pull out of it without rereading it. Open format and cross-linking between notes are the two conditions; the tool isn't one of them.
What this changes for an organization
Past decisions, meeting notes, procedures, customer feedback: internal knowledge stops being a dead warehouse and becomes a queryable asset, without the data ever leaving the company's perimeter. It's the foundation for any serious internal use of AI — sovereign, durable, independent of whichever tool is fashionable this year. And it's a structural project before it's a software one: the choice of format and the filing order decide everything AI can do afterward.
This is the foundation I set up at Orogen, from what I've built and measured on my own base. Write to me, let's talk.
Sources
- Andrej Karpathy, LLM Wiki, note published on GitHub, 2026 — the three-layer model: raw sources, compiled wiki, schema — gist.github.com.
- Measurement of the MoonLab base as of September 13, 2026: 8,934 Markdown files, 1,565 synthesis notes, 211 state-of-the-art reviews, 18 top-level folders — count performed on the repository itself.
- On the silent failure of the nightly automation: see the resource The AI that dreamed.