Give the Brain Nouns

Research status. Typed relationships and graph retrieval exist in the systems I use and study. A complete operational ontology across my second brain remains future work.

My notes contain many versions of the same thing. A repository can appear as a URL, a project name, a meeting topic, and a line in a task list. A person may appear by name in one document and by role in another. Retrieval can find these mentions. It still has to decide whether they refer to the same entity.

An ontology makes that decision explicit. It gives the brain nouns.

Instead of treating every mention as fresh text, the system can recognise an Article, Project, Person, Claim, or Source. Each object has properties, and each relationship has a name. An article cites a source. A claim appears in an article. A project owns a repository.

A folder tells me where something was filed. An ontology tells me what the thing is and how it relates to the rest of the system.

That difference matters because real work crosses folders. This essay belongs to a research track, continues a previous article, draws on several sources, tests a design idea, and may later become part of a talk. Forcing it into one location hides most of those relationships.

A graph can preserve them without making seven copies of the article. It also gives the agent a controlled vocabulary. When I ask for “unfinished second-brain writing supported by CodeGraph research”, the agent can traverse named relationships instead of hoping those words happen to occur together.

The most useful objects carry state. An Article can be draft, reviewed, or published. A Claim can be supported, disputed, or future work. A Source can have a retrieval date and a freshness rule.

Those states should not be inferred from whichever paragraph an agent retrieved first. They should be read from an authoritative object, with the document retained as evidence.

The temptation is to model everything. That creates an elegant schema nobody can maintain. I prefer to start with the nouns that change what the agent should do.

For a personal publishing system, five may be enough: Article, Claim, Source, Series, and Repository. The useful relationships are equally small: supports, continues, implements, and contradicts.

The test is practical. Does adding the type prevent a real mistake, improve a useful query, or make authority clearer? If it does none of those, it may only be decorative structure.

Named relationships still need provenance. Some links are asserted by me. Some are extracted from documents. Some are suggestions awaiting review. The system should not flatten those into the same level of truth.

This is especially important when personal and company knowledge remain separate. A shared name does not grant permission to join records across those boundaries. The ontology must respect ownership before it improves convenience.

Once the nouns and relationships are stable, the next question becomes possible: what can the system reliably do with them?

• Previous: Documents Are Evidence, Not the World
• Next: Give the Brain Verbs
• Related: From Company Knowledge to Verified Changes