Knowledge Graph, Second Brain or Plain Search: What Is a Company Brain Made Of?

People come to this from three directions. Someone who reads about productivity has heard of a second brain. Someone technical has heard that the answer is a knowledge graph, possibly in Neo4j. And someone practical asks why the business cannot just search the shared drive.

All three are describing parts of the same thing. Here is what a company brain is actually made of, in plain terms, and when each part is worth the effort.

Layer one: documents

The foundation is written pages that say how the business works. How a job is done. What each system holds. What was decided and why. Who owns what.

This is the part most businesses skip, and no technology fixes it. Search over documents that were never written finds nothing. A knowledge graph of pages that are out of date is a well-organised record of how things used to work.

For most businesses, the documents are plain text: markdown files, a Notion space, or a Google Drive folder. Plain text matters because every AI tool can read it. Claude, ChatGPT, Copilot and Gemini can all work from the same pages, and changing tools changes nothing in the brain.

What to put in a company brain first lists the twelve pages to start with.

Layer two: search and retrieval

Once the pages exist, something has to find the right one when a question comes in. That is search, and with AI tools it is usually called retrieval: the tool looks through the pages, pulls out the ones that look relevant, and answers from them.

Search is good at one thing: finding pages that use the words in the question, or words close in meaning. Ask "what is our refund policy?" and it will find pages that mention refunds.

What plain search cannot do well:

• Tell which page is current. If there are three pages about refunds, an old one, a draft and the real one, search returns all three. An AI tool may answer from the wrong one.
• Follow a chain. "Who do I ask about a refund over the usual limit?" needs the refund page, the page that says who decides exceptions, and that person's page. Search finds pages one at a time; it does not know they are connected.
• Know what the words mean here. Every business has terms that mean something specific. Search treats them like any other word.

This is why "just search the shared drive" works for finding a file you know exists, and fails for answering how the business works.

Layer three: links, which make it a graph

A knowledge graph sounds technical, but the idea is simple. It is a set of things and the relationships between them. Neo4j, the best known graph database, describes it in the same terms: nodes are the things, relationships connect them, and both can carry properties.

In a company brain, the things are pages, and the relationships are links. The workflow page for "monthly invoicing" links to the system page for the accounting software, the decision page that set the payment terms, and the person page for whoever owns it. Follow the links and you get the whole picture, not one page at a time.

Two habits turn a pile of linked pages into something that behaves like a proper graph:

Every page has a kind. Workflow, system, decision, person. When a page declares what it is, a link from a workflow to a system means "this job uses this tool", and a link from a decision to a person means "this person made the call". The links carry meaning.

Links are deliberate. A workflow page names its systems, its owner and the decisions behind it, every time. That consistency is what lets a person, or an AI tool, start at any page and find its way to the answer.

With those two habits, a folder of linked markdown files is a knowledge graph in every way that matters for a business. Each page is a node, each link is a relationship, and the kinds give them meaning. An AI tool can read the page it needs, follow its links and come back with an answer that accounts for the decision behind it and the person who owns it. What is a company brain shows the page kinds in more detail.

When a graph database is worth it

A graph database such as Neo4j stores nodes and relationships natively and lets you query them with a dedicated query language (Neo4j's is called Cypher). It is a serious tool, and there are jobs it does far better than linked pages.

It earns its place when:

• The graph is large. Many thousands of nodes, often generated from systems rather than written by people: every customer, order, product and supplier, and how they relate.
• The questions are about relationships. "Which customers bought products from suppliers that had a quality issue last quarter?" is a graph query. Linked pages cannot answer it quickly.
• Several systems feed it. The CRM, the accounting system and the operations software all contribute, and the graph is where they meet.
• Someone will maintain it. A graph database needs a data model, loading jobs and a person who understands both.

It is the wrong first step when:

• The knowledge is a few hundred pages that people write and edit by hand.
• Nobody on the team will own a database.
• The main readers are people and AI chat tools, which read pages far more easily than they query a database.

For most small and mid-sized businesses the order is clear. Write the pages. Link them, with page kinds. Connect them to the AI tools the team uses. Add a graph database later, if and when you have questions about relationships across systems that the pages cannot answer.

Putting it together

A working company brain uses all three layers:

• Documents hold the knowledge.
• Links with page kinds turn the documents into a knowledge graph, so a question can follow the chain from a job to the system, the decision and the person behind it.
• Search finds the starting page.

Add one more thing that no technology supplies: an owner and a regular check for pages that have gone stale or contradict each other. Without that, any of the three layers drifts out of date.

On our second brain pages we describe the result as how your business works, written down once as a knowledge graph, so every person and every AI tool reads the same thing. It is also the context layer under the AI Operating System, the method we use to set up a team's AI tools: the skills and agents read the brain before they start a job, and write back what they learned when they finish.