You can build a working company brain for yourself this weekend. Set up an Obsidian vault, write your definitions and unwritten rules as one note per fact with an owner and a date, then connect your AI over MCP so it answers from your notes instead of guessing. For one person it works well. Then you try to roll it out to your team, and it cracks: nobody keeps contributing, every connected tool sees every note, conflicting notes carry equal authority, and stale notes get quoted with confidence. A company brain for a whole company is a different system. That is what Sento builds.
Published on: July 6, 2026·10 min read
A hands-on guide to writing your company down so AI can read it. We start with tools you already have, and we are honest about where that stops working.
Here is how to build a company brain, in five steps: set up an Obsidian vault, write your definitions as one note per fact, capture the unwritten rules the same way, connect your AI to the vault over MCP, and keep it honest with a weekly review. By Sunday night your AI can answer questions about your business from your own notes instead of guessing. This guide walks every step, with the exact starter notes and the config to copy.
A company brain is a single, current source of how your business works, readable by both people and AI tools. The full definition, and why Y Combinator named it one of the ideas it most wants built in 2026, is in What Is a Company Brain?. This is the practical half.
One honest note before we start. The weekend version works, and you should build it, because nothing teaches you the shape of the problem faster. And somewhere around the moment you try to roll it out to your team, you will discover why the DIY version stays personal. We will be specific about that too.
Step 1: Set up the vault
Download Obsidian, create a vault called something like company-brain, and give it a small folder structure. Resist the urge to design a taxonomy for an hour. Four folders cover almost everything:
Definitions/ for what your words mean. Rules/ for how things actually get decided. Playbooks/ for repeatable processes. Facts/ for the specific truths that fit nowhere else, like which account is on non-standard terms.
The reason Obsidian and not a Google Doc: Obsidian stores everything as plain Markdown files on your disk. That matters later, because plain files are the one format every AI tool can read without an importer. You are not locked in, and the brain is machine-readable from day one.
Step 2: Write definitions as one note per fact
The single most important craft decision: one note, one fact. Not a page called "Sales stuff" with forty things on it. A note called Active customer.md that says exactly one thing:
An active customer is an account with a paid contract AND at least one login in the last 30 days. Finance counts paid contracts only; do not use their number for product decisions. Owner: Lena. Last checked: 2026-07-04.
Three things to copy from that example. It states the definition in one sentence a machine can quote. It names the known conflict (finance counts differently) instead of pretending the company agrees. And it carries an owner and a date, because a definition without those is a rumor.
Write ten of these to start, no more. If you want a starter set, these ten cover the disagreements almost every company has, steal the list and adjust:
Active customer
Qualified lead
ARR (and what it excludes)
Churn (logo or revenue, and as of when)
Trial conversion (what counts as converted)
The renewal approval rule
The discount policy (who can give what)
Refund policy (the real one, with the exceptions)
Which accounts are on non-standard terms, and why
The two or three accounts that are sensitive, and why
The test for whether a term belongs on your list: does it get two different answers in the same meeting at your company. If yes, it goes in the vault first.
Step 3: Capture the unwritten rules
This is the part that makes it a company brain rather than a glossary. The rules that run the business are mostly not written anywhere: how renewals actually get approved, why a certain account gets a discount, which customers are sensitive and why.
Two techniques work. The first is Slack archaeology: search your own history for phrases like "just so you know," "for context," and "the reason we," and you will find dozens of rules that were explained once in a DM and never written down. The second is the exit-interview question asked early: sit a colleague down for twenty minutes and ask "what do you know that would hurt us if you forgot it tomorrow." Write each answer as its own note in Rules/, same format, owner and date included.
Link notes to each other as you go with Obsidian's [[double brackets]]. [[Acme]] links to [[Net-60 exception]] links to [[Renewal approval]]. The graph is not decoration; it is how the next reader, human or machine, finds the context around a fact.
Step 4: Connect your AI to it
Now the payoff. Obsidian vaults are folders of Markdown, so you can serve them to an AI tool over MCP, the open protocol AI tools use to read external context. The simplest route is Claude Desktop plus the reference filesystem MCP server, pointed read-only at your vault. Open Claude Desktop's settings, edit claude_desktop_config.json, and add:
Swap in your vault's path, restart Claude Desktop, and it can read every note. That is the whole setup: one config block and a restart. If you outgrow plain file access, community-built Obsidian MCP servers add proper search over the vault, but start simple; for a few hundred notes, file access is enough.
Then ask your AI a real question you would normally answer from memory: "What are the renewal terms for Acme?" Watch it answer from your note instead of guessing. That moment is the entire thesis of this category in miniature. The model did not get smarter. It just finally had something true to read.
Add a weekly fifteen-minute review to keep it honest: scan the notes touched by real questions that week, fix anything stale, add the two or three facts you explained out loud to someone. That ritual is the maintenance cost of the personal version, and for one person it is genuinely sustainable.
What you now have, and for whom it works
For an individual founder, operator, or consultant, this setup is quietly excellent. Your AI answers from your actual definitions. New chats stop starting from zero. You have, at personal scale, exactly what the fancy vendors describe: written-down context, served over MCP, with owners and dates.
We mean it when we say build this. Everything you learn writing your first thirty notes, which facts matter, where the conflicts hide, how fast things go stale, is knowledge you will need no matter what you run later.
Why the DIY company brain doesn't scale to a company
Now try to give it to your team, and watch four specific things happen.
Rollout is a behavior change nobody signed up for. The vault worked because you, one motivated person, wrote the notes. Now you need the sales lead, the ops manager, and the finance person to write Markdown in a shared vault with consistent formats, owners, and dates. Obsidian's sync and collaboration options exist, but you are effectively asking your whole company to adopt a note-taking tool and a librarian's discipline at the same time. In practice one champion maintains it, the champion gets busy, and contribution stops in week three.
Security is all or nothing. A filesystem MCP server pointed at the vault gives the AI, and anyone whose tool connects, everything. The salary discussion note sits next to the glossary. There is no way to say the support bot may read product policies but not board notes, or that a contractor's Claude sees definitions but not customer terms. At personal scale that is fine because it is all yours. At company scale it is the reason your security-minded colleague vetoes the whole thing, and they are right to.
There is no canonical. In your solo vault, every note is true because you wrote it. In a shared vault, two people write conflicting notes about the discount policy and both sit there with equal authority. Nothing marks one as reviewed and current and the other as someone's draft opinion. The AI reads both with the same confidence, which brings back the exact three-answers problem you built the brain to kill.
Staleness stops being visible. Your weekly fifteen-minute review does not survive contact with 40 contributors and 600 notes. Notes rot silently, and an AI quoting a rotted note is worse than no brain at all, because it is wrong with authority. There is also no trace: when the AI gives a wrong answer, nothing tells you which note it read, so you cannot even find what to fix.
None of these are Obsidian's fault. It is a personal knowledge tool doing its job well. The gap is that a company brain at company scale is not a bigger vault. It is a different kind of system: one that fills itself from the tools the company already uses instead of waiting for humans to write notes, that has a review step so canonical and draft are different things, that serves each AI tool only what the asker is allowed to see, and that logs what every answer was based on.
Side by side, the two problems look like this:
The company-grade version
That system is what we are building at Sento. It seeds the first draft of your company brain automatically from the warehouse, dbt, Notion, and Slack, so nobody authors from a blank page and it does not go stale when the champion gets busy. Non-data teammates add the unwritten rules through a simple capture flow with a review queue, so the data team decides what becomes canonical. Every AI tool your company runs reads from it over MCP, filtered by who is asking, and every answer traces back to the exact fact it used. A two-person team stands it up in about a week, and it is free during early access. The architecture term for all this is a context layer, and the deeper story of the unwritten rules is in Company Knowledge for AI Agents.
The honest summary: build the Obsidian version for yourself, this week, because it works and it will teach you the problem. When the question becomes "how does the whole company run on this," that is a different problem, and it is the one we spend our days on.
Your AI doesn't know your company. Sento fixes that.
Can I really build a company brain in Obsidian? For yourself, yes, and this guide shows the full setup: one note per fact with an owner and a date, unwritten rules captured as notes, and an MCP connection so your AI reads the vault. For a company, the vault model breaks on rollout, access control, canonical-versus-draft, and staleness, which is why the org-grade version is a different kind of system.
What about Notion or Confluence instead of Obsidian? The same pattern works and the same limits apply. Wiki tools are built for people reading, so once a whole team depends on them they hit the vault's same walls: no per-asker access filtering for AI, no authority model, silent staleness. They make excellent seed sources for a governed system, which is exactly how Sento uses them.
Do I need to be technical to set up the MCP connection? Mildly. Connecting Claude Desktop to a folder via an MCP filesystem server is a config file, not a coding project. If you can install Obsidian plugins, you can do this in an evening.
What is actually different about Sento versus a well-run shared vault? Four things the vault cannot do: it self-seeds and refreshes from the company's existing tools rather than depending on people writing notes, a review queue makes canonical and draft different things, the MCP Server filters what each AI tool sees by who is asking, and Lineage and Audit traces every answer back to the fact it used. The vault is a personal brain. This is the company's.
