Files · Obsidian + git
Where knowledge is written
Plain files your team opens and corrects. They belong to you, keep a full history and work with any model. Agents propose changes; a person approves them.
Knowledge architecture · Three stages · Never rebuilt
One company brain. It grows with you, and is never rebuilt.
Every AI agent is only as good as what it knows about your business. We build that knowledge once, as files you own, and add layers as you add agents: an index when there is more to search, permissions when different people need different answers. Nothing you wrote is ever migrated.
What is an AI company brain, and how does it scale?
An AI company brain is the shared knowledge every AI agent in a business reads from. theagency47 builds it in three layers: plain markdown files in an Obsidian vault as the source of truth, a search index (PostgreSQL with pgvector, or Cloudflare Vectorize and D1) that agents query, and live connections through MCP to systems such as your CRM, so live data is never copied. It scales in three stages: files for one agent, files plus an index for several agents, then permissions, memory and logs for many agents and audiences. The files never move, so each stage adds a layer without a rebuild.
theagency47 · Updated October 2026 What we deliver
People write in files. Agents search an index. Live data stays where it lives.
The full architecture, as it runs at Stage 3. Stages 1 and 2 are the same picture with fewer layers switched on.
Knowledge flow Data flowProposed change, approved by a personLive via MCP
Three layers, three jobs
Each layer does one thing well.
Most knowledge projects fail when one tool tries to be the editor, the search engine and the system of record at once.
Index · PostgreSQL + pgvector
Where agents search
Built from the files and rebuilt on every change. If it were lost tomorrow, it would be back in minutes, because the source never left the vault.
Live systems · MCP
Where live data stays
Bookings, invoices and stock are never copied. The agent asks your CRM, ERP or booking system at the moment it needs the answer, read-only wherever possible.
Grows without being rebuilt
Three stages. Each one adds a layer. Nothing is migrated.
You start with the smallest brain that does the job. When the business adds agents, sources or audiences, we add the next layer on top of the same files.
Files
One agent, one team · 1 agent
Obsidian vault · markdown + git built
- How agents search
- The agent reads a vault index and opens the right files. No database needed, and more accurate at this size.
- What this stage adds
- Your knowledge, structured: entities, processes, rules, decisions, FAQ. Owner and review date on every file.
- Move up when
- More than a few hundred pages, or a second agent.
Files + index
Several agents, shared knowledge · 3 agents
Search index · PostgreSQL + pgvector added
Obsidian vault · markdown + git unchanged
- How agents search
- Hybrid search (keywords plus meaning) across the vault and large sources such as tickets, emails and contracts.
- What this stage adds
- An index rebuilt automatically from the files on every change, plus test questions that re-run after each sync.
- Move up when
- A customer-facing agent next to internal ones, or a second department.
Files + index + control
Many agents, many audiences · An agent per department
Permissions · memory · logs added
Search index · PostgreSQL + pgvector unchanged
Obsidian vault · markdown + git unchanged
- How agents search
- Each agent searches only what its audience may see: public, internal or confidential. Structured questions run as SQL.
- What this stage adds
- Per-audience access, per-agent memory, a log of what every agent read and did, and cost limits per team.
- Move up when
- This is the full architecture. You grow by adding agents and sources.
What stays exactly the same at every stage
- Your files and folders. The vault you wrote at Stage 1 is the vault at Stage 3.
- Owners and review dates. The same people stay responsible for the same knowledge.
- Your agents’ instructions. They point to the same files; only the way they search gets faster.
- The test questions. The same 20–30 questions with known answers prove each stage works before it goes live.
- Ownership. Every file, index and configuration is yours, and goes with you.
Every stage includes OS47, the workspace where your team sees what your agents know and do, approves what they propose before anything is sent, and checks their test scores. See OS47 →
When to move up
Six signals that it is time for the next stage.
Until one of these happens, the smaller stage is the better one. None of them is a reason to start bigger than you need.
| Signal | Stage | Why it matters |
|---|---|---|
| The vault passes a few hundred pages | 1 → 2 | Reading files one by one gets slow and misses things; an index finds the right passage in milliseconds. |
| A second or third agent reads the same knowledge | 1 → 2 | One index keeps every agent answering from the same version of the truth. |
| Answers must come from thousands of tickets, emails or contracts | 1 → 2 | Large raw sources are indexed alongside the curated vault, not pasted into it. |
| A customer-facing agent joins internal ones | 2 → 3 | The chatbot must be unable to reach margins, salaries or internal notes. That needs permissions, not instructions. |
| A second department gets its own agents | 2 → 3 | Each department sees its own knowledge plus the shared layer, with its own logs and budget. |
| You need to show what an agent read and why it answered | 2 → 3 | Per-agent logs and memory turn every answer into an auditable trail. |
Many departments
Company brain at enterprise scale
In a large company the brain is the same three layers, with more people writing to it and more agents reading from it. Four things change at Stage 3:
- An owner per knowledge area. Finance owns the finance files, HR owns the HR files. Every file has a named owner and a review date, so no central team has to keep everything current.
- Permissions per audience. Public, internal and confidential knowledge are separated in the index, so a customer-facing agent cannot reach what only staff may see, and one department's agents cannot read another's confidential notes. See how client and audience isolation works.
- A log per agent. What each agent searched, read and did is recorded, for the agent's owner to review and for audit.
- One brain, not one per department. New departments add folders and sources to the same vault and index. Nothing is copied, so a price or a policy is changed once.
The brain is the foundation every phase of an AI transformation builds on. Who decides which agent may see what is set out in our AI governance framework.
Where it runs
Your cloud, ours, or Cloudflare.
The architecture is the same wherever it runs. Most businesses start with the vault on their own computers and a managed PostgreSQL in an EU region, such as Supabase. Businesses that want no servers at all run it on Cloudflare: the index becomes Vectorize and D1, the agents run on the Cloudflare Agents SDK, and access goes through Cloudflare Access. See where OS47 runs.
Whichever you choose, the rules on our trust page apply: a person approves what agents change, every action is logged, and your data is never used to train models.
Keeping it alive
Brains rarely fail on technology. They fail on upkeep.
- Agents propose, people approve. No agent writes to the vault directly; every change waits for the file’s owner.
- An interviewer agent fills the gaps. Two or three questions a day to the team, so knowledge gets written down without a project.
- Stale files surface on their own. Anything not reviewed in 90 days goes to its owner once a month.
- Save to brain in one step. A good answer in chat or a meeting becomes a proposed file change with one command.
- Tests on every change. If the score on the known-answer questions drops, the change is rolled back.
FAQ
Questions about the company brain.
What is an AI company brain?
An AI company brain is the knowledge layer every AI agent in a business reads from: your processes, rules, prices, decisions and answers to common questions, written down once and shared by all agents. theagency47 builds it in three layers: plain markdown files in an Obsidian vault as the source of truth, a search index (PostgreSQL with pgvector, or Cloudflare Vectorize and D1) that agents query, and live connections to systems such as your CRM or booking platform through MCP, so live data is never copied.
Should a company brain live in Obsidian or in a database?
Both, in different roles. Obsidian is where people write and correct knowledge: plain files you own, readable by any model, with a full change history. A database such as PostgreSQL with pgvector is where agents search once there are many documents, many agents or different audiences that must see different things. The files stay the source; the database is an index rebuilt from them, never the other way round.
Do I have to rebuild the brain when the business grows?
No. Stage 1 is the file vault. Stage 2 adds a search index built from the same files. Stage 3 adds per-audience permissions, agent memory and logs on top of that index. Your files, folder structure, owners and test questions stay exactly as they are at every stage; each step adds a layer and nothing is migrated.
When should I move from Stage 1 to Stage 2?
Usually when one of three things happens: the vault grows past a few hundred pages, a second or third agent starts reading from it, or answers need to come from thousands of emails, tickets or contracts rather than curated pages. Until then an agent reading the vault index directly is simpler and more accurate than a vector database.
When do I need Stage 3?
When agents serve different audiences. The moment a customer-facing chatbot sits next to internal agents, the brain needs permissions so the chatbot can never reach salaries, margins or internal notes. Stage 3 also adds per-agent memory and a full log of what each agent read and did, which multi-department setups and audits require.
Is live data like bookings or invoices copied into the brain?
No. Anything that changes by the minute stays in the system that owns it: CRM, ERP, property-management or booking system, invoicing. Agents query it live through MCP connections at the moment they need it. The brain holds only the knowledge that exists nowhere else, the knowledge that today lives in your team’s heads.
How much does it cost?
Stage 1 is the Second Brain build: a fixed €4,500, delivered in 3–4 weeks, with €1,500 off when bundled with a Workforce or Scale package build. Stages 2 and 3 are added as fixed-price upgrades scoped in the AI Audit (€297, credited if you start within 30 days), because their size depends on your document volume, number of agents and audiences.
