An article on the Open Knowledge Format makes a case I recognise. Stop asking a vector search to guess your company’s definition of revenue. Keep it as plain Markdown files, linked explicitly, in git. An agent follows links instead of hoping the right chunk comes back. Then comes the part that worries me: background agents keep those files up to date, fixing links as code changes.
I keep my own notes this way, and agree with the first half. Files you can diff are worth more than embeddings you cannot read. But an agent that edits your corporate truth is a risk of its own. Every change goes through a pull request, and a check that needs no model runs first: does every link resolve, and did a human sign off on the facts that matter. Skip that gate and you have swapped a guessing database for a guessing librarian.