How to build a company brain, part 9: when it gets big and slow, give the agent a graph

Three things grow with the brain. Search gets slower, because scanning every folder for a word takes longer each month. Context gets heavier: the agent reads whole files into its context window, so the input tokens per question go up, and so does the bill. And the blind spots multiply. An agent can't follow a link in a file it never opened, so the more files there are, the more it misses.

I fix this in two steps. The first is the index in part 10, which finds files fast. The second is a graph. It tells the agent which few files matter for a question and how they connect.

Graphify is a tool that turns a folder into a knowledge graph. You install it once, type /graphify in your coding assistant (its README lists Claude Code, Cursor, Codex and Gemini CLI, among others), and get three files back: the graph itself, a page you can open in a browser and click through, and a report in plain language. The report names the most connected nodes, which Graphify calls god nodes, and connections you might not expect.

The graph holds nodes and the links between them. Every link carries a tag that says whether it was read directly from the source or inferred. Closely linked nodes are grouped into communities and given names. There are also hyperedges: one edge that groups three or more nodes, such as all the steps of one workflow.

Graphify was built for code, and that shows in how it treats the two kinds of input. Code is parsed on your machine with tree-sitter. No AI model is involved, so the result is the same every time and costs nothing. Documents are different. Markdown, PDFs and images go through what the README calls a semantic pass, which your assistant's model (or an API key you configure) runs. That pass costs tokens, and it means the model reads those documents. Keep that in mind before you point it at a folder with personal data. Part 7 covers what stays out.

Once the graph exists, the agent asks it before reading files. There are three commands. "graphify query" takes a question and returns a small piece of the graph. "graphify path" shows how two things connect. "graphify explain" describes one node and its neighbours. The agent gets back a handful of notes and links, and opens only those.

A business brain is mostly documents. For Graphify to be any use there, the notes have to carry their own links. This is what I do:

• Link generously. Every note links to what it depends on with a relative Markdown link, and lists its source ids in the header. A feature brief links the feedback it came from. A decision links the analysis behind it.
• Store the links. The index in part 10 keeps every Markdown link in a table. That table is already a simple graph: it knows who links to whom, which notes nobody links to, and which pages everyone cites.
• Let Graphify add meaning. Its pass over the Markdown adds typed edges such as "references", "conceptually related to" and "rationale for", plus hyperedges for workflows.
• Keep it fresh. Running "graphify hook install" adds a git hook that rebuilds the graph after every commit. For notes there is also an update mode that re-extracts only the files that changed.

In one of my brains the graph has about 3,000 nodes and about 9,900 links. More than 7,200 of those links are "references" edges that come from notes. Fewer than 100 are "calls" edges from code. So the shape of the business sits in its documents, and the links are what make that shape visible to an agent.

Two lists are worth reading every week. The orphan list lists notes nobody links to, which usually means a missing summary or a misfiled document. The hub list shows which pages many others cite, so those are the pages that have to stay accurate. At Kranich Studio, the example company in this series, the hubs would be the bakery chain's client page, the waitlist knowledge page and BRAIN.md. When a hub changes, the pages that cite it get a review.

I haven't measured how many tokens the graph saves per question in my own brains, so I won't give you a number. Graphify's README links to its own benchmarks if you want one. Measure on your own folder before you trust a published figure.

Goal: a weekly health check of this brain's knowledge graph.
Steps: update the graph with Graphify (changed files only). Then read graph.json and our links table.
Report: the ten most-cited pages and the date each was last verified. Every note with no incoming links, with a suggestion for where to link it. Any page cited by more than five others that changed this week.
Constraints: do not edit, move or delete any note. Give me the list and nothing else.
From now on: for a question that crosses folders, query the graph first and read only the files it returns.