A company brain is a persistent, permissioned memory of what an organization knows — documents, messages, meetings, records, decisions — structured and kept current so that people can query it and AI agents can act on it. What a company brain is covers the definition and the category; this guide covers how to build a company brain, in the order that has worked. It draws on the build write-ups published in 2026 by practitioners who did it — a YC founder running a startup from a git repository that learns, an agency operator giving every agent shared context, an investor's prototype-to-production sequence, an infrastructure vendor's workflow-first method — and on the Karpathy pattern of an LLM-maintained wiki that many of them use. They disagree on tooling and agree on sequence: one workflow first, the sources it needs, a memory the agents maintain, and expansion only after the first version pays for itself.
Three Decisions Before You Start
Who is the primary reader? If the answer is "employees searching," build an AI knowledge base and stop; it is cheaper. A company brain is justified when the primary reader is an agent that has to act — because agents cannot ask a colleague, and they will act on whatever the brain holds.
What is the first workflow? Not the first department. A workflow: renewal outreach, inbound lead research, weekly pipeline review, customer-onboarding, incident triage. The brain's first job is to make one set of agents good at one thing.
Who owns truth? When the CRM and a Slack thread disagree, which wins, and who decides? Encode this before ingestion, because the brain will surface the disagreement on day one.
How to Build a Company Brain, Step by Step
- Pick the workflow and the agents. Choose the one with the highest pain and the clearest outcome metric. Name the two or three agents that will run it — research, drafting, execution — and the human who reviews their output.
- Collect five real examples. Take five recent instances of the workflow done well by a person and write down every piece of context they used: which accounts, what history, which exceptions, which documents, who they asked. This is the brain's initial specification. It is also where tribal knowledge gets surfaced, because the person will name things that exist nowhere.
- Connect only those sources. The CRM, the call recordings, the shared drive folder, the two Slack channels, the ticket queue — whatever the five examples touched. Continuous sync, not a one-time export. Breadth comes later, and most brains that fail failed here by connecting everything.
- Choose the memory architecture. Three patterns cover almost every 2026 build; they are compared below. The choice depends on team size, how technical the builders are, and whether the brain must be a system of record.
- Extract entities, decisions and time. Whatever the architecture, the brain has to know the nouns the workflow reasons about — accounts, contacts, products, competitors, deals — the decisions and exceptions that govern them, and which version of each fact applies now. A vector index alone does not do this; an entity layer or context graph does.
- Enforce permissions at the source. The brain inherits who may see what from the systems it ingests from, and it filters before an agent or person sees anything. Retrofitting permissions is the most expensive mistake on the list.
- Give agents read and write. Agents retrieve context at runtime and record what they did — the email sent, the account flagged, the decision taken and why. This is the loop that agentic memory describes, and it is what stops the brain's knowledge ending at the last human-authored document.
- Measure and expand. Score the brain on the workflow's outcome — meetings booked, renewals saved, hours removed, errors avoided — and on answer accuracy against a fixed question set. Add the next workflow only when the first one is producing. Every practitioner account of a brain that works describes this shape; every account of one that did not describes the opposite.
Three Architecture Patterns
Pattern | How it works | Best for | Watch out for |
|---|---|---|---|
Agent-maintained wiki (the Karpathy pattern) | Raw sources in one folder, a structured wiki in another, a schema describing how the wiki is organized; an LLM agent ingests new material, updates pages, resolves contradictions and answers questions. Often a git repository, so every change is versioned and reviewable | Small teams, technical founders, one to three agents, documents and text as the main sources | Scale: hundreds of sources and structured data strain a text wiki; permissions are coarse |
Retrieval index plus entity layer | Documents chunked and embedded into a vector index with hybrid search; an entity and decision graph extracted alongside; agents query both and write outcomes back | Mid-size companies with mixed sources and several workflows; teams with engineering capacity | Maintenance: connectors, evaluation and freshness are ongoing work |
Managed platform | A company-brain or agent-memory product provides ingestion, entity model, permissions, agent interfaces and write-back | Companies that want production in weeks, non-technical operators, regulated environments | Lock-in and data residency; verify that the entity layer and write-back are real, not a RAG demo |
The first pattern is where most 2026 builds began, because it can be running in a weekend with a CLI agent and a repository. The write-ups of a startup run from a git repo that learns, and of an "organisational brain" taken from prototype to production, both start there. The second is what the first becomes when it outgrows text. The third is what teams adopt when the brain has to be a system of record for agents acting on revenue, and the products are compared in company brain software.
What a Company Brain Costs to Build
The first pattern costs a few days of a technical person's time plus model usage. The second costs an engineer for the connectors, evaluation and entity extraction, and ongoing time to keep it right — typically one part-time owner. The third is a platform license plus usage. In every case the dominant cost is not the technology; it is the discipline of scoping narrowly and the work of deciding who owns truth. Budget for a first version in weeks and a company-wide brain over quarters.
Mistakes That Sink Company Brains
- Connecting everything first. Produces a search box over a swamp. Connect what the first workflow needs.
- Treating retrieval as memory. RAG fetches chunks. A brain also knows entities, decisions and what changed. Without the second half, agents act on stale facts with confidence.
- Read-only agents. If agents cannot write back, organizational memory stops at the last human document and the brain decays.
- Permissions later. They never come later. Inherit them from the source on day one.
- Measuring documents indexed. The metric is decisions made and outcomes produced. A brain with a thousand sources and no measurable outcome is a cost.
- No owner of truth. The brain will find contradictions. Someone has to have the authority to resolve them.
A Worked Example: The Revenue Brain
The narrow brain most established companies should build first is about revenue. The workflow is "know how buyers in our category find and choose, and act on it." The agents are three: one that maps demand — every question buyers ask ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, and which competitors get recommended; one that models which moves would grow revenue; one that builds the assets and takes the actions. The sources are AI answer engines, the website, the CRM and the pipeline. The memory is a proprietary database the company owns, written to by all three agents every cycle, so that the brain knows not only what buyers asked but what the company did in response and what moved the number. That is the system MultiplierAI's Recon, Strategist and Closer agents run, and it follows every step above: one workflow, the sources it needs, entities and decisions over time, agents that write back, and a metric — revenue proven — that the brain is judged on.
Frequently Asked Questions
What is a company brain?
A persistent, permissioned, continuously updated memory of what an organization knows, structured so that people can query it and AI agents can use it to act correctly. It differs from a knowledge base in that its primary reader is an agent, it holds entities and decisions over time, and agents write to it.
What is Y Combinator's company brain?
Y Combinator's Summer 2026 Request for Startups asked founders to build "an AI-native knowledge system that continuously ingests everything a company produces" and makes it queryable by every employee — with tests such as answering decision-history questions with citations and respecting source permissions. It named the category.
How long does it take to build a company brain?
A narrow one — one workflow, a few agents, the sources they need — can be running in days on the wiki pattern or weeks on a platform. A company-wide brain is a multi-quarter program that grows outward from a narrow version that already works.
Can I build a company brain with Claude or ChatGPT?
Yes, for the first pattern: a CLI agent such as Claude Code maintaining a structured wiki in a repository is how many 2026 builds started. What the model does not provide is ingestion connectors, an entity model, permissions and write-back at scale — those are the build or the platform.
What is the difference between a company brain and RAG?
RAG is a retrieval technique the brain uses. A company brain adds continuous ingestion, an entity and decision model, versioning of facts, permissions, and agent write-back. A RAG demo over a document folder is not a brain.
References
- https://www.ycombinator.com/rfs
- https://vectorize.io/articles/how-to-build-company-brain
- https://elevationcapital.com/perspectives/ai-organisational-brain-plotline
- https://thenewaiorder.substack.com/p/i-built-a-company-brain-to-run-my
- https://x.com/ericosiu/article/2060415100603781497
- https://www.askglitch.com/blog/build-your-company-brain
- https://github.com/garrytan/gbrain
- https://www.reddit.com/r/AIMemory/comments/1taordw/how_to_build_a_company_brain/