Most businesses do not have a knowledge problem. They have a location problem. The knowledge exists, it is just distributed across four inboxes, a chat tool, a shared drive nobody has tidied since 2023, and the head of whoever has been there longest. Everything still works, right up until that person is on holiday, or leaves, or you try to hand a client to someone new.
A second brain fixes that only if it stops being a place you file things and becomes the place work is defined. This piece is about what that looks like in practice, what it genuinely replaces, and the parts where it will let you down if you oversell it to yourself.
The shift from filing to operating
The mechanic is simple. You keep one governed set of Markdown files: raw sources that are never edited, a maintained wiki the agent writes, and a small constitution the agent reads on every task. The structure and the reasoning behind it are covered in why I built Lorekeep, and the original pattern is Karpathy’s LLM wiki.
The shift happens when those files stop describing the work and start carrying it. A process is not a document you wrote once, it is the thing the agent follows when you ask it to run that process. A client is not a folder, it is an entity with linked decisions, commitments and history. The knowledge base becomes the interface to the operation rather than a record of it, and that is the whole difference.
What it actually holds
Nine things, in rough order of how quickly each pays for itself.
- Processes, written so a machine can run them. Not a policy document. A sequence with the assumptions stated and the exceptions named. The test is whether an agent can execute it without you in the room, which is also the test of whether a new employee could.
- Handovers. Every process page names its owner and what a handover needs. The first time someone takes over a workstream from a page rather than from a two-hour call, the system has paid for itself.
- Client context. What was agreed, what was promised, what was tried, what failed, and who said so. Linked to the raw source, so the answer to any argument is a transcript rather than a memory.
- Decisions and their reasons. The single highest-value thing in the vault and the one almost nobody keeps. A decision page records what was chosen, what it replaced and why. Six months later this is the difference between changing course and relitigating.
- Brand and tone rules. Once the style guide is canon, every document the agent drafts follows it by default. Consistency stops being something you police in review and becomes something you get for free.
- Task tracking that knows the context. Because tasks live beside the projects and people they belong to, the agent can generate a personal board for each person from the same source, rather than everyone maintaining their own list in their own tool.
- Relationships between work. A wiki is a graph. That is not decoration: it is how you see that four projects depend on one integration, or that a client owns three unfinished threads.
- The dull operational furniture. Holiday capacity, who covers what, onboarding checklists, supplier terms, renewal dates. None of it is exciting and all of it currently lives in somebody’s calendar.
- Money, if you choose to put it there. Run rate, utilisation, what a retainer assumes, which clients are actually profitable. This one is a deliberate decision, not a default, and I come back to it below.
What the owner gets
Three things, and the first is the one people underestimate.
Visibility without asking. If the work is defined in the vault, the state of the business is a query, not a status meeting. That is a different relationship with your own company: you stop being the bottleneck through which information travels and start being able to check.
Consistency you do not have to enforce. Every output starts from the same rules, so the floor rises. The gap between your best work and your fastest work narrows, which is usually where reputations are actually made or lost.
Time back at the start of everything. The unglamorous saving is the ten minutes of context you re-explain at the start of every task, every session, every briefing. It is small each time and enormous in aggregate, and it is the first thing you notice.
What it displaces, and what it costs
Be careful here, because this is where these projects get oversold. A folder of Markdown files does not replace a system of record. It replaces the layer of context those systems never held.
In practice it tends to absorb the lightweight tools bought to hold documents, notes, wikis and internal process, the ones with a per-seat price and a low usage rate. It does not replace accounting, payroll, a CRM with a pipeline your team actually works, or anything with a legal obligation attached. Those keep their systems; the vault holds the reasoning around them.
The real cost is not licence fees, it is discipline. Someone has to feed sources in and accept or reject the pages that come back. That is a much smaller job than writing a wiki, which is why this generation of the idea survives where the last one died, but it is not zero, and a vault nobody feeds is a filing cabinet with nothing in it.
Why model agnostic matters
Everything is plain text in ordinary folders, under version control. No database, no proprietary format, no vendor holding your operating knowledge as a hostage to their pricing. Any capable agent can be pointed at it, and when a better one arrives in eight months you point it at the same files.
It also stays on your machine, which is why it can hold commercial terms and client work without a policy conversation first. Sync is a choice you make rather than a feature imposed on you: a private remote if you want the same knowledge on two machines, an encrypted drive if you would rather it never touched a third party, nothing at all if that suits. Make it deliberately. A knowledge base worth building is a knowledge base worth protecting, and the version of this system that holds your finances is not the version to sync casually.
Where it stops
Four honest limits.
- It is not a permissions system. A folder is a folder. If different people must see different things, you need separate vaults or a real system with access control.
- It does not fix an operation that is not thought through. Writing a process down exposes that it was never agreed. That is valuable and it is also work, and no amount of automation does it for you.
- It decays without a sweep. Claims that were true when captured go stale silently. Anything machine-checkable should be re-checked automatically; anything positional needs a date and a review cadence. Without that, the vault ages into confident fiction.
- It is not judgement. It makes better decisions possible by making the context available. It does not make them.
How to start on Monday
Do not begin with a taxonomy. Begin with one painful, person-bound process, the one where everyone says ask Sarah. Capture the raw sources for it. Let the agent draft the page. Accept the page. Then have someone who is not Sarah run the process from it and see where it breaks.
That single loop teaches you more than a month of structure. If it works, do the next one. The step by step build is in how to build your own second brain, and the governance I run on top of it is packaged in Lorekeep, which is free and open source.
By Piers Butler, founder of Laurelin Labs. This is the system behind the frameworks and the client work.