Everyone's building a "company brain." Nobody solved the personal one first.

In the last two weeks the same idea showed up three times, in three different rooms, from people who don't seem to know about each other. A team-collaboration platform launched with the tagline "the company brain your AI runs on," letting teams share agents, credentials, and knowledge without exposing secrets to each individual tool. A separate open-source project shipped a self-hosted vector database so several major AI tools can all read from the same persistent memory of a person's work, no SaaS middleman required. On a developer forum, someone asked a plain question that got seven replies in a day: what is your company brain setup?

Three answers to a question nobody had to ask two years ago. How does a team give its AI tools a shared, durable memory of what the company knows?

The pitch is easy to follow. A support agent that can't see last week's incident postmortem starts from zero. A coding agent that doesn't know your team already tried and rejected a particular architecture will propose it again next sprint. Every AI tool a company adopts arrives with amnesia, and "company brain" is the current name for the cure.

This is not a new problem. It's the personal knowledge management problem, argued over in note-taking forums since long before anyone pointed an LLM at it. Give a system a place to put everything you know, and two questions show up. How much structure do you impose up front? Who, or what, decides where a new piece of information goes?

Right now most company-brain efforts answer both questions the same way: top-down. A widely shared post this week lays out "4 essential files" every AI workspace needs: a company file, a customer file, an offer file, a voice file. Decide the categories first, then pour the business into them. It's the same instinct behind that memory project's schema and the collaboration platform's shared-context model. Define the buckets, then fill them.

Steph Ango has argued against that instinct for years, just at the scale of one person instead of one company. His guide to how he uses Obsidian describes something closer to the opposite: almost no folders, heavy internal linking between notes, and a daily journal treated as the entry point for everything. Structure emerges from use instead of getting imposed before he's written a word. He calls the result fractal: the same rough shape repeats whether you're looking at a single day's journal or five years of notes. Nobody hands Ango a template with four required files on day one. The organization shows up later, built out of links he made because two ideas were genuinely related, not because a form told him to file them next to each other.

The bottom-up approach has a track record the top-down one doesn't have yet. Ango's system has held up across years of daily use. The company-brain products are weeks old. And the failure mode of top-down categorization is familiar to anyone who's used a shared drive at a real company: the categories fit for six months, then the business changes, and half of what's filed under "offer" turns out to describe a customer segment that didn't exist when the folder was made. A rigid schema doesn't bend. It calcifies. People start dumping files wherever, and the schema turns into decoration.

None of this makes top-down wrong or bottom-up an automatic winner. A completely unstructured pile of notes fails too, just quietly: the AI can technically read all of it and still can't find the one thing that matters, because nothing marks it as important or explains why it was kept. The real answer, at company scale or at the scale of one person, probably needs both. Light structure that doesn't fight how people actually think, plus enough signal about intent and importance that retrieval doesn't turn into a haystack search.

That second piece, capturing why something was saved and not just what it is, is what almost nobody building a company brain is thinking about yet, because they're still solving the more basic problem of getting data into one place at all. It's also the problem anyone building tools for personal reading and bookmarking has to solve first, at a much smaller scale, before one person's saved articles and half-finished notes are worth handing to an agent. Get that right for a single reading pile, tagged by why something mattered and not just what domain it came from, and you've effectively prototyped the harder version of the problem a company brain is trying to solve for forty people at once. That's the layer I keep coming back to when I think about what Saive should be doing with a save beyond just storing the URL.

The teams racing to build shared AI memory right now are running the same experiment individual note-takers have run for a decade: how much structure, imposed by whom, makes a pile of information useful later. The personal side has a ten-year head start. Company brain builders would do well to read it before they ship version two.