Organisational Memory in the Age of AI: Building the Company Brain

Earlier this year the series argued that 2026 would be the year of agents. Over the summer, a Forbes article suggested something else might matter more: the company brain. Many agents have not worked as well as promised, and a large part of the reason is that they have no memory of the business they are working in. So Rob Woodhead asks Rob Lynch and Ben Burch what a company brain actually is, why connecting AI to a shared drive is not the same thing, and whether a business should start building one now. Their answer is that the hard part is not the technology. It is working out what your business actually knows, where that knowledge lives, and who should be allowed to see it.

Key takeaways

• A company brain is a structured version of what your business knows, built so that AI can read it. It exists because chat windows run out of context and rereading every document for every question does not scale.
• Connecting AI to SharePoint or Google Drive is not the same thing. Without a map of what matters, an abandoned document called final process can outweigh the one people actually follow.
• The most valuable knowledge usually lives in people’s heads, so the map starts with interviews rather than an upload.
• Documents are snapshots and the business keeps moving. A useful brain needs a feedback loop and a human in the loop for meaningful updates.
• Permissions have to be designed source by source: structured timelines can be shared widely, email should be opt-in, and HR material stays restricted.
• Answering questions is the tip of the iceberg. Agents built on a reliable brain can keep dashboards current and alert the right person before a deadline slips.
• Start small and start now. Mapping your data sources on paper is useful whatever the technology does next, and the text files you produce will carry over to new tools.

Episode chapters

• 00:00 Welcome back after the summer break
• 01:15 From the year of agents to the company brain
• 02:06 What a company brain actually is
• 03:14 Context windows and why long AI chats lose the thread
• 06:31 Why connecting AI to your drive is not enough
• 10:00 Documents go stale: timelines and feedback loops
• 12:18 Where to start: building a map of the business
• 16:54 Asking the same question through different lenses
• 20:25 Permissions and sensitive data
• 24:21 Agents built on top of the company brain
• 29:02 Beyond questions: dashboards, schedules and alerts
• 31:39 Start now or wait six months?
• 35:24 The offline groundwork that pays off later
• 39:05 You do not need an all-singing, all-dancing brain
• 40:06 Wrap-up

A company brain is context, not a longer chat window

A company brain is a version of your company’s knowledge that AI can actually read and make sense of.

Rob Lynch starts with the simplest version. Condense what the business knows, as far as that is possible, into something a language model can understand. The obvious first use is a company chatbot that can genuinely answer questions, whether you joined last week or have been there for a decade. Everything else builds out from there, but the hard part is the same throughout: getting the data in, taming it and making it usable.

Ben explains why it is needed by describing something most people have already run into. A long conversation with an AI assistant starts well and then drifts. Answers get weaker, earlier details go missing and repetition creeps in. That is the context window at work, the limited span of text a model can draw on at once. It is not really memory, and it was never designed to hold a whole business.

The brute-force alternative, rereading every document for every question, does not scale even for a small company. Ben’s comparison is reading your entire inbox every time you want to answer one question. A company brain organises the material so that the useful conversations become easy: a new starter asking the process questions they would rather not put to their line manager, or someone sifting old proposals for what actually won.

Giving AI access to your drive is not the same as giving it context

Most companies could connect an AI tool to all of their files tomorrow. Rob Lynch’s point is that this alone gives the AI a badly distorted picture.

Give Claude access to Google Drive, or Copilot access to SharePoint, and nothing tells it which documents matter. A file from 2021 called final process for XYZ, which nobody has followed in years, can carry far more weight than it deserves. The process people actually use might sit in a simpler document somewhere else, or not be written down at all and live only in people’s heads.

So the work is in context. The AI needs a map of what is important and what is not, sitting as a layer between the question and the data. That map ultimately comes from people. Do it well and the company brain earns its keep. Skip it and you do not get hallucinations so much as answers that are misattributed and wrong-headed.

If you are working out what that map could look like for your own business, that is the sort of question we tend to start with. You can read about how we work or see the products we have built.

Business knowledge goes stale, so the brain needs a feedback loop

Documents are snapshots. The business is not.

Ben’s concern is the timeline. A document reflects what was known when it was written, while the business and the market around it keep moving. Newer material usually supersedes older material, but what was true yesterday may not be true today, and a single rule change can make a long email trail obsolete overnight.

He describes the validity of information as multi-dimensional. A junior’s opinion in a reply to the CEO is not the same kind of evidence as a decision the CEO has made, and a useful brain needs to tell the difference. Without that, you have built a glorified search engine, and as Ben puts it, we have had search for ages. Finding information faster is not intelligence. The opportunity is to be considerably smarter than that.

The map of the business comes from the people who run it

Old.St Labs is building a company brain internally, and Rob Lynch is candid that nobody has settled on the right way to do it yet.

The direction is clear, he says, even if the method will look different by 2027 or 2028. The approach emerging as best practice for now starts with people. Interviews with the CEO, product managers and others across the business about the most relevant projects, documents and processes are valuable even when the output is unstructured, because they come from the people who actually know the answers.

Those interviews become a set of text files describing the key elements of the business: its projects, its people, its processes. Each file points to the documents that genuinely matter for that area. When something new arrives, such as an email about a project, it can be read against those files and the map updated.

Any meaningful update keeps a human in the loop. Rob’s example is a weekly check-in where the system tells the CEO what it thinks has changed and asks whether that is right, over Teams, Slack or email. The CEO talks it through, and the extra context goes back into the files. Even then the map will drift, so a plan for keeping it current has to be built in from the start.

The value is in asking the same question through different lenses

AI tools are very good at doing the same task over and over again, quickly. Ben thinks that is where much of the value of a company brain sits.

Take a set of past proposals, each marked as won or lost. When they were written, everyone thought they were the best proposal possible. With a company brain you can keep going back over them, comparing them against something that launched yesterday and re-ranking what mattered as new information arrives.

It is close to how human memory works, Ben argues, and just as incomplete. Someone who has been in a business for ten or twenty years brings a sense of its trajectory that a snapshot of today cannot. Founder and CEO insight tends to matter most, because they know how the business got where it is. Because the outputs are kept as readable text, the brain can be questioned again from a different angle at any point, and that human filtering is what it is trying to replicate.

Permissions have to be designed one data source at a time

Every business holds information that not everyone should see. Rob Woodhead’s first example is pay.

Rob Lynch’s answer is that each data source needs to be considered on its own terms. Structured data, such as a set of project timelines, sits at one end. It can usually be shared across the company, and it is less likely to go stale as long as the timelines themselves are maintained.

Email sits at the other end. It is easy to misread out of context and some of it will be personal, so it needs a completely different approach. One option is opt-in: only messages a person labels, or files in a particular folder, go into the shared brain. People could also keep a private layer of their own for project information that is not for general view, and something like an HR manager’s discussions needs to be in the brain without going out to the whole company.

None of this is exotic. It starts with listing your data sources, deciding which parts of each you will ingest and how, and setting the context layer so the AI can make good use of what it is given. Done properly, it also answers the question everyone is asking at the moment: where is my data going, and how will it be used?

Agents need a reliable brain before they can do much with it

Agents are not dead, Ben says. They are just getting started.

Not long ago, the industry conversation was all about RAG databases, and setting up an agent was a technical job. Now it can be done with clicks and typing. None of the earlier ideas have gone away; the tools have simply become powerful and accessible enough for many more people to try things. When Ben started on the company brain idea a year ago, the useful comparison was a team of graduates you could send to dig around under strict rules. Now it is closer to a team that has been in the business for ten years, and it needs far less precise guidance to give the same quality of answer.

Rob Lynch sees question-answering as the tip of the iceberg. His favourite example is project management. An email arrives that would normally need a graduate, or a long-serving colleague, to read, interpret and turn into a schedule update. With a reliable brain underneath, that can happen automatically. Dashboards stay current with the qualitative state of a project, not just the financial numbers, and agents can watch them, flag when something is approaching a threshold or a deadline, and ping the right person.

All of it depends on the brain itself: its robustness, consistency, quality and fidelity.

If you are weighing up where agents could fit into your own operations, you can see how we approach AI solutions.

The groundwork is worth doing now, whatever the technology does next

Rob Woodhead closes on the question a business owner would ask: start now, or wait six months for the industry to settle?

Ben’s answer is to try it. A pilot might take a few weeks of elapsed time, but only hours of actual work. Pick something helpful and uncontentious, see whether there is value in it, then talk to people who have done it before about scaling it up or working out what went wrong. He has little time for posts showing AI getting a question wrong. Ask the same question now and it will often get it right, and humans get things wrong too. That is not a reason to stop checking the output, but it is not a reason to avoid experimenting either. If you tried something six months ago, try it again.

Rob Lynch adds that much of the real work is not technical at all. Sketching your data sources on paper, how they flow and who should see what, will be just as useful in six months even if you have not touched the technology. Anything you do build will be text files describing your business. They will need updating, but they can be moved to a new system as the tools change.

Nor does it have to be all-singing, all-dancing. If someone senior is buried in enquiry emails, a simple brain built from a handful of documents can draft replies from your own data. Ben’s final point is the human one: talk to the people who are trying this. It is changing all the time, and comparing notes is still the quickest way to understand it.

About the hosts

Rob Woodhead is a co-founder of Old.St Labs, a London-based software agency building web and mobile products for startups and innovators, and hosts the podcast.

Rob Lynch is a co-founder of Old.St Labs. Throughout the episode he draws on the company’s work building its own internal company brain.

Ben Burch is a Technical Consultant at Old.St Labs and a regular on the series, having appeared in the earlier episodes on vibe coding, agents, and AI data and security. He began working on the company brain idea around a year ago.

Frequently asked questions

What is a company brain?

A company brain is a structured version of a company’s knowledge that AI can read and understand. Rob Lynch describes the simplest use as a company chatbot that any employee can ask about the business and get a reliable answer, with agents and automation built on top.

Why is connecting AI to SharePoint or Google Drive not enough?

Access alone does not tell the AI which documents matter. Rob Lynch’s example is an outdated final process document being given more weight than the process people actually follow, which may not be written down at all. A company brain adds a map of what is important, drawn from the people who know the business.

Why do AI chats get worse the longer they go on?

Language models work within a context window, a limited span of text they can draw on at once. As a conversation grows, earlier details fall away and answers weaken. Ben Burch explains that a company brain is meant to move that context out of individual chats and into an organised structure.

How do you keep a company brain up to date?

Build the upkeep in from the start. Ben Burch describes it as a living process with a continual feedback loop, because newer information supersedes older information. Rob Lynch suggests keeping a human in the loop, such as a weekly check-in where the CEO confirms what has changed.

How do you handle sensitive data in a company brain?

Treat each data source separately. Rob Lynch suggests sharing structured data such as project timelines widely, making email opt-in so people choose what joins the shared brain, and restricting sensitive areas such as HR to the people who need them.

Do AI agents still matter if you have a company brain?

Yes, and they become more useful. Rob Lynch sees the brain as the foundation for agents that keep project dashboards up to date and alert the right person when a deadline or threshold is close. How well that works depends on the quality of the brain underneath.

Should a business start building a company brain now?

Ben Burch recommends a small, low-risk pilot to see whether there is value. Rob Lynch adds that offline work, such as mapping your data sources and permissions, stays useful whatever happens to the technology, and that a simple brain built from a few documents can solve a specific problem like answering enquiry emails.

‍