The standard way to connect an AI agent to a body of knowledge is retrieval: chop the documents into chunks, index them, and fetch the most relevant pieces when a question arrives. It works, as far as it goes. The agent gets the right paragraph.
What it doesn’t get is everything an expert knows about that paragraph. Which claims are load-bearing and which are speculative. Which terms mean something narrower than they appear to. What order to consult things in. Where the knowledge runs out. Every practitioner carries reading instructions like these — a methodology — and almost none of it survives the trip into a vector database. The agent is handed a library and no librarian.
Doxa is a bet on fixing that at the format level. A .doxa file is a single SQLite database holding structured knowledge — concepts with summaries, confidence scores, provenance, bibliographic sources, and a typed graph of relationships between ideas. All of that is data, and useful. But one table is different in kind: _guidance. It holds instructions, not content — the methodology, query strategy, interpretation rules, and known limitations of this specific body of knowledge, written by the person who built it. A system prompt for the knowledge base, authored by its expert.
The premise underneath is one the whole field keeps re-learning: an AI system’s output quality is set by the context available at the moment of generation. Retrieval supplies the content half of that context. The guidance table supplies the judgment half — and because it lives inside the file, it arrives wherever the file goes. Any agent that knows the doxa convention reads the guidance first and then navigates the knowledge the way its author would: same priorities, same cautions, same sense of where the edges are. That’s the design intent, anyway — designed to make a generic agent behave less like a search engine and more like someone who has worked in the domain.
Doxa began as a practical problem, not a theory: making a long-running personal research corpus portable — knowledge an agent could be taught to read rather than merely search. The format generalized from there. The pieces now exist as an open spec, a universal agent skill that learns any doxa’s domain from its guidance table at runtime, and a server that lets any compatible agent query a doxa over the network. The spec and tooling are tested. The format also tries to be honest about what knowledge is: the name comes from the Greek for practiced judgment — what Plato ranked below certified truth. That ranking is the point. A curated knowledge base can’t honestly ship certified truth; it can ship an expert’s judgment with confidence scores, provenance, and a changelog.
What this doesn’t show. The load-bearing question hasn’t been tested. The real test of the thesis is a cold read: hand an agent that has never seen a given doxa the file alone — no coaching, no custom prompt — and check whether the guidance layer actually makes it perform like a domain specialist rather than a chunk-fetcher. That test hasn’t run. Until it does, “teaches agents how to read” is a design claim, not a demonstrated capability. There’s a second honest limit: only one substantial doxa exists so far, and it was built by the same person who wrote the spec. A format proves itself on the second author and the second domain, not the first. Both tests are on the near-term list; we’ll publish what happens, including if the answer is no.
If your organization sits on hard-won knowledge that your current tools treat as a pile of documents, the Brief is where that conversation starts.