The Company Brain
Plain-text operational memory that every AI tool inherits on day one.
What it is
A structured knowledge base holding what the business actually knows: client history, supplier behavior, pricing methods, and the reasoning behind past decisions. Three layers, with source material at the bottom, distilled pages in the middle, and a navigation layer on top that keeps retrieval cheap.
How it works
Every work session, email sweep, and research pass writes back into it, so it compounds instead of decaying. A retrieval server indexes the whole thing and answers questions over it, locally and from a cloud twin, including by voice.
Development to deployment
This is the least visible system I run and the highest leverage one by a distance. Everything else on this page works because this sits underneath it. When a tool changes or a person leaves, the knowledge stays.
Value delivered
Over six hundred operational pages and growing every day. Every new tool I adopt starts with years of context instead of a blank prompt, which is why builds that should take a week take an afternoon.
The problem is not that you lack documents
Most small businesses have plenty of documentation. It is in a shared drive nobody opens, a wiki that stopped being true in 2023, forty thousand emails, and the head of the person who has been there longest. The information exists. It is simply not retrievable at the moment a decision needs it.
That becomes expensive in three specific ways. New hires take months to become useful because the context is undocumented. Decisions get re litigated because nobody remembers why the last one went the way it did. And every AI tool you adopt starts from zero, so you spend the first twenty minutes of every session explaining your own business to it again.
The company brain fixes the third one first, and the first two fall out as a side effect. It is the single highest leverage build I run, and it is the one clients are least likely to ask for because it does not demo well.
Three layers, and why the middle one is the product
The bottom layer is raw. Immutable source material: transcripts, email dumps, exported reports, research. It is never edited, only added to. That immutability is what makes the whole thing trustworthy, because you can always trace a claim back to where it came from.
The middle layer is the wiki. Distilled pages, one per entity, that hold what is currently true about a client, a supplier, a product, a process, or a person. This layer is rewritten constantly as reality changes. It is the layer that answers questions, and it is the actual product.
The top layer is navigation. A small set of index and recency files that let a retrieval system find the right page without reading everything. This is the layer that keeps queries cheap. Skip it and your retrieval costs scale with the size of the knowledge base, which is exactly the wrong direction.
Why plain text beats a database here
The whole thing is markdown files in folders. No proprietary format, no vendor, no export process. That decision gets questioned every time I explain it and it has never once been wrong.
Plain text means every tool can read it. The retrieval server, the local editor, a language model, a shell script, a future tool that does not exist yet. It means version control gives you a full history of what the business believed and when. It means there is no migration when you change tools, because the knowledge was never trapped in the tool.
It also means the failure mode is graceful. If every piece of software around it disappeared tomorrow, you would still have a legible, organized, human readable account of how the business works. That is not a small thing to be able to say about a system.
It has to write itself or it dies
Every knowledge base project I have watched fail died the same way: it depended on humans remembering to update it. Discipline is not a system. Within a quarter the pages are stale, people stop trusting them, and stopping trusting them is the same as deleting them.
So capture is automated and continuous. Work sessions write back into it. The daily email sweep files thread outcomes onto the relevant client and supplier pages. Message and channel activity gets routed to the entities it touches. The human contribution is correction and judgment, not transcription.
That inversion is the entire trick. The system captures by default and interrupts a person only when something genuinely needs a decision. Compounding beats discipline, and it is the only version of this that is still alive after a year.
What it costs and what it returns
The infrastructure cost is close to nothing. Files on disk, a retrieval index, and a cloud mirror. Across my entire operating stack, every AI system I run, the all in monthly bill sits comfortably under three hundred dollars, and the knowledge layer is a rounding error inside that.
The return is that new builds start with years of context instead of a blank prompt. Work that should take a week takes an afternoon, not because the coding got faster but because the specification was already written down. When someone asks what we agreed with a client eighteen months ago, that is a question with an answer now.
The uncomfortable part is that this build is worth the most and photographs the worst. It is folders of text. There is no dashboard to show anyone. I am telling you about it anyway because a business that installs this before it installs anything else gets more out of everything that follows.
Common questions
What is a company brain, in plain terms?
A structured, continuously updated knowledge base holding what your business actually knows: client history, supplier behavior, how you price, and the reasoning behind past decisions. It is built so both people and AI tools can retrieve from it, which means every new tool you adopt starts already knowing your business.
How is this different from a Notion wiki or a shared drive?
Structurally, by the separation of immutable source material from distilled current truth, and by a navigation layer that keeps retrieval cheap. Operationally, by the fact that it writes itself. A wiki that depends on people remembering to update it is stale within a quarter, and that is why most of them are.
Do we need AI for this to be worth it?
No. The organizational value stands on its own, and onboarding alone usually justifies it. AI retrieval is what turns it from a filing system into something that answers questions, so it is where most of the leverage shows up.
How long does it take to build one?
The structure goes up in days. Getting real context into it is the part that takes actual time and depends entirely on what shape your existing material is in. I would rather look at yours and tell you honestly than quote a number now that we both know is invented.
What does it cost to run?
Very little. It is files plus a retrieval index. My entire AI operating stack, of which this is one part, runs under three hundred dollars a month all in. The expensive input is the thinking about what belongs in it, not the infrastructure.
Get the full build doc
The deep documentation on my builds, architecture, guardrails, and lessons included, lives in the client portal. A free account gets you in, and the docs land in your files.