The second brain a business runs on.

The idea is older than the apps.

Niklas Luhmann, a German sociologist, published more than seventy books and hundreds of articles in his career. He attributed the output to a slip-box, a Zettelkasten he started building in the 1950s that grew to roughly ninety thousand index cards. Each card held one idea, written in his own words, and pointed to other cards. The value was not the cards. It was the links between them, which let connections surface that he had not planned. The system was a thinking partner, not a filing cabinet.

The modern version carries different names and the same shape. Tiago Forte packaged it as Building a Second Brain, with two pieces worth keeping: PARA, a way to sort everything into Projects, Areas, Resources, and Archives, and CODE, a verb sequence for what you actually do with a note (Capture, Organize, Distill, Express). Andy Matuschak pushed on the note itself with the idea of evergreen notes: write each one to be atomic, concept-oriented, densely linked, and addressed to your future self rather than an audience. Strip the branding off all three and the same instruction remains. Capture in your own words, link aggressively, and let the structure emerge from the links instead of imposing it up front.

Why the company wiki rots.

Most companies already pay for a place to put what they know. Notion, Confluence, or SharePoint sits there, and it rots anyway. The reason is structural, not a failure of discipline. A wiki is built to be read by everyone, so writing in it is a small act of publishing, and publishing has friction. The half-formed thought that would have been worth keeping never gets written down, because writing it cleanly enough for the team to read costs more than the thought feels worth in the moment. So the wiki fills with stale onboarding pages and dead links, and the actual operating knowledge stays where it always was, in email threads, in chat scrollback, and in the founders' heads.

The cost of that gap is well documented. A McKinsey analysis clocked the average knowledge worker at close to a fifth of the workweek spent searching for and gathering information, a number anyone who has hunted for the current version of a file across three apps already recognizes. The deeper problem is not the time spent looking for what was written down. It is everything that never gets written down at all.

Recent estimates put roughly seventy to eighty percent of enterprise knowledge as tacit, held in people's heads and never recorded (Gartner, 2024), while the annual cost of knowledge-worker turnover in the United States runs near 1.3 trillion dollars, with the lost institutional context, not the price of rehiring, as the largest share (Deloitte, 2024). Average knowledge-worker tenure now sits around four years. A wiki holds the slice of knowledge someone judged worth publishing; the tacit majority leaves on a four-year clock. A second brain does not fix this by being a better wiki. It fixes it by lowering the cost of capture, so the tacit majority has somewhere to land before it walks, and by separating the two jobs a wiki tries to do at once: the private place where thinking accumulates, and the shared place where settled knowledge is published. Collapse them and you get neither.

Why Obsidian, specifically.

Obsidian is a free desktop and mobile app that treats a folder of plain-text Markdown files as the database. There is no proprietary store. The notes are .md files on disk, readable in any text editor, greppable from the command line, and version-controllable with git. The app's stated philosophy, from its CEO, is file over app: the artifact that should outlast the software is the file, not the program that happens to open it. That single property is why a business should care. A SaaS wiki owns your knowledge and rents it back to you; if the vendor changes terms or shuts down, your knowledge is hostage to an export feature. A Markdown vault is yours in a format that will still open in twenty years.

The features that matter for thinking are the linking ones. Type [[note name]] and you have a wikilink; the target note automatically gains a backlink showing every note that points to it, which is the digital version of Luhmann's card pointers. A local graph view shows the neighborhood around a note. The Dataview plugin, and the newer built-in Bases feature, let you query notes by their properties, so a folder of meeting notes can render itself as a live table without manual maintenance. None of this requires a server or a subscription.

Two costs are real and worth stating plainly. The first is money, and it is small: the app is free for personal use, Obsidian Sync for encrypted multi-device sync runs $4 per user per month billed annually (or $5 month to month), and Obsidian Publish, if you ever want to put part of the vault on the public web, is $8 per month per site. The second cost is discipline. The plugin ecosystem is large and seductive, and a vault with thirty plugins is slower, more fragile, and more likely to break on an update than one with five. The working set that covers most needs is small: a templating plugin, Dataview, daily notes, and a tag manager. Treat every additional plugin as a liability to justify, not a feature to collect.

A starter vault you can build today.

Structure is where most attempts die, because people design an elaborate taxonomy before they have any notes to file. Resist that. A vault that works for an operating business needs six folders and nothing more at the start:

• 00_Daily. One note per day, created automatically. Everything starts here: the call you just got off, the idea in the shower, the thing a client said. Capture has to be frictionless or it does not happen.
• 01_Concepts. Crystallized, wikilink-rich notes, one idea each. This is where a daily capture goes when it earns a permanent home.
• 02_Reading. Notes on books, articles, talks, and research, in your own words. Reference material that compounds.
• 03_Promote. A holding pen for notes that are ready to leave the vault for a more public or canonical home. The weekly ritual empties it.
• _Archive. Old daily notes and abandoned drafts. Searchable, out of the way.
• _templates. The skeletons for daily notes, concept notes, and reading notes, so each new note starts consistent.

Layer a small set of status tags on top, so a note's maturity is visible at a glance: #raw for an unprocessed capture, #cooking for something being worked, #promote for ready-to-migrate, #promoted once it has left, and #archived. Five tags, not fifty. The point of the structure is to make the next action obvious, not to model the universe.

The ritual, and the boundary that keeps it alive.

A capture system without a processing habit becomes a landfill. The habit is one block of time a week, the same slot every week. Pick Sunday evening or Monday morning, it does not matter which, only that it recurs. In that block you do four things: scan everything tagged #promote, decide each one's real home and move the content there, retag the vault copy #promoted with a link to where it landed, and sweep daily notes older than ninety days into the archive. Half an hour, most weeks. If you run the ritual four weeks in a row it sticks and the vault stays useful. Skip it for a month and the vault becomes noise you stop opening.

The boundary is the rule that separates a second brain from a junk drawer. State it as a direction of flow: the vault is the inbox, the company's canonical systems are the outbox. The vault is where thinking starts and where half-formed things are allowed to be messy. When a thought crystallizes into something the business will act on, it leaves: into the project folder, the contract, the ticket tracker, the published document. What does not belong in the vault is as important as what does. Client operational records, financial and billing data, signed contracts, and anything resembling a credential or a secret have canonical homes elsewhere and stay there. Passwords live in a password manager. API tokens live in a secrets store. A note that asks "where would I look for this later" answers its own filing question; if the answer is "the contract folder," it does not go in the vault.

The 2026 turn: the vault an agent can read.

For a decade the second brain was a tool for one human's thinking. That is no longer the whole story. The format the method has always used, plain Markdown in a folder, happens to be the format an AI agent handles best. A model reads a Markdown file with no parsing layer, no API, and no export step. The wikilinks are legible structure. The folder names are context. The status tags tell the agent what is settled and what is still raw. A knowledge vault is, almost by accident, the cleanest possible substrate for an agent to operate over.

The connective tissue is no longer hypothetical. Anthropic's Model Context Protocol is an open way to connect agents to a company's data and tools, and a folder of Markdown notes is one of the simplest things to expose through it. As of 2026 there are several open-source MCP servers built specifically to give an agent read and write access to an Obsidian vault. The filesystem-first ones read the Markdown straight off disk, need no plugin, and work whether the app is open or not; others run through Obsidian's local REST API to reach the app's own features. Either way the same vault now serves two readers. A person searches it and links in it; an agent reads it for context at the start of a task and writes structured notes back into it.

Context is the variable that moves the result. An agent without the company's accumulated knowledge guesses; an agent that can read the vault works from what the company actually knows. One vendor analysis measured a thirty-eight percent improvement in query accuracy when a model was given governed institutional context instead of left to infer it. The discipline that makes a vault useful to a human, atomic notes, honest tags, a real boundary, is the same discipline that makes it safe and productive for an agent to act on. A messy vault produces a confused agent for exactly the reasons it produces a confused colleague.

Pointing an agent at the vault is a real decision, not a free upgrade. Whatever the agent reads becomes part of the model provider's context, so the boundary that keeps client records, financials, and secrets out of the vault is the same boundary that makes the vault safe to expose. A vault disciplined enough to hand to an agent is one that never had a password in it to begin with. That is the through-line of the whole method: the rule that keeps a second brain useful to a person is the rule that makes it usable by the systems built on top of it.

How Rarefied Earth thinks about this work.

The firm runs exactly the setup described above, and it runs it as two separate substrates on purpose. There is a thinking-substrate vault, an Obsidian vault of Markdown notes synced across desktop and mobile, and there is the canonical operating workspace where client work, billing, contracts, and published artifacts live. The vault is the inbox; the operating workspace is the outbox. Capture is allowed to be messy in the vault precisely because the weekly ritual migrates anything that crystallizes out to its canonical home and the boundary keeps client data, financial records, and secrets from ever landing there.

Keeping the two separate is the part most teams get wrong. The temptation is to let the low-friction capture surface absorb everything, at which point it quietly becomes a second operating system competing with the real one, and the company now has two half-maintained sources of truth instead of one good one. One inbox, one outbox, one direction of flow. That is the whole trick, and it is the same trick that lets an agent read the vault for context without tripping over a contract or a password that was never supposed to be there.

The posture is the one that runs through everything the firm builds: the unglamorous infrastructure layer is the part that actually compounds. A second brain is not a productivity hack. It is the memory a company stops rebuilding from scratch every time someone forgets, leaves, or hands work to an agent.

Sources and further reading.

Related work.

A second brain is the personal-scale version of a larger argument the firm makes about infrastructure. The field guide on why most company AI projects stall covers the shared-memory layer agents need to be useful in production, and shows where a knowledge vault fits in the wider operating substrate a company actually runs on.