Method

Cite it, or say you don't know.

"Knowledge is of two kinds. We know a subject ourselves, or we know where we can find information upon it."Samuel Johnson, 1775

Every company runs on two kinds of knowledge. The first lives in documents: the handbook, the wiki, the runbooks nobody updates. The second lives in people: how the discount really gets decided, which customer can be pushed and which cannot, why the deploy waits until Thursday. The second kind is most of what a company actually is.

The industry's answer has been to ask people to write more things down. It has never worked, because the people who hold the most judgment have the least time, and because judgment does not survive being flattened into a wiki page. Documentation describes the system. It rarely describes the judgment.

Then AI arrived and made the gap dangerous. An agent reads everything you wrote and knows nothing you didn't. Worse: when it doesn't know, it guesses, in fluent confident prose. A guess in citation-shaped formatting is more dangerous than no answer at all. One wrong confident answer about a pricing exception, a compliance boundary, or a customer promise costs more than a thousand right ones save.

The three rules

We hold that an organizational memory must earn trust, and that trust has exactly three requirements.

  1. Provenance, always. Every answer names who knew it, where it came from, and when someone last confirmed it was still true. Knowledge you can check is knowledge you can act on.
  2. Refusal, when weak. When the grounding isn't there, the system says "I don't know" in plain words. The refusal is the product. It is what makes the citations worth believing.
  3. Elicitation, then memory. A refusal is not a dead end. Athena asks the one person who knows, captures the answer at the moment it is given, and keeps it for everyone who comes after. The knowledge gets written down at the only moment anyone is willing to: when someone needs it.

Why now

Because agents are starting to act. For thirty years, stale knowledge meant a colleague gave you an outdated answer and you found out in the meeting. From now on, stale knowledge means an autonomous system executes it. The tacit layer was an annoyance in the age of documents. In the age of agents it is the bottleneck, and the companies that wire it up first will compound in a way the others cannot copy by buying tools.

We are building Athena with ten design partners: teams small enough to move, senior enough to know where their knowledge dies. If the rules above read like something you already believe, you are probably one of them.

David & Erik Two technical founders · St. Gallen & Zürich