Documents Are Evidence, Not the World

Research status. Sourced retrieval and separate personal and company memory are working now. The operational model described here is the next layer I am testing.

Imagine asking an agent how to publish an article. It finds a clear checklist, a previous release note, and a conversation where I changed the process. All three documents are relevant. Only one may describe the process I use today.

The agent has found evidence. It has not yet found the current state of the world.

This is the limit I keep reaching with retrieval. Search can locate a procedure, but it does not automatically know whether that procedure is still binding. It can find a project plan, but it does not know whether the project is active. It can quote a decision without knowing whether I later reversed it.

Documents preserve explanations. They hold the reasoning behind a decision, the wording of a policy, the details of a meeting, and the evidence behind a claim. That history matters because an agent should be able to show where an answer came from.

My current second-brain work treats provenance as part of the answer. A useful result should identify its source, its age, and the boundary around what was searched. The same principle appears in my CodeGraph work: a dependency result carries evidence and limitations rather than pretending to be a complete safety guarantee.

Documents answer a vital question: What has been recorded about this?

A document should not be forced to represent live operational state. A sentence saying that an article is ready for publication is different from the article actually having the state ready. A meeting note naming an owner is different from the system currently assigning that person responsibility.

If these are collapsed into one thing, the agent starts treating prose as authority. That is dangerous. A draft can look like a policy. A historical incident can look current. A suggestion can look approved.

The answer is not to replace documents with a database. The two layers do different jobs.

The operational layer can say that an article is a draft, belongs to the second-brain series, cites four sources, and is waiting for my review. The document layer can explain the argument, preserve the source material, and show why I made each editorial choice.

A strong agent should move between them. It can inspect the current article object, retrieve the notes supporting a disputed claim, and return with both the state and the evidence. If the two disagree, the disagreement becomes visible instead of being smoothed over by a fluent answer.

Today, I have durable memory behind multiple agent harnesses, explicit routing between personal and company knowledge, sourced pages, and graph-aware retrieval. That is already more useful than a pile of unconnected chat histories.

I do not yet have one finished public system where every retrieved document resolves cleanly to a live operational object. That connection is the research programme.

The next step is to give the brain nouns: explicit things such as Article, Claim, Source, Project, and Decision, with relationships that an agent can inspect instead of infer again on every turn.

• Previous: The Last Project You Build Alone
• Next: Give the Brain Nouns
• Related: Why a Knowledge Management Layer Still Matters