If you're about to build your company a brain, or pay someone to build one, read this first.
I built my company a brain. The first thing it did was lie to me about how smart we are.
A vault is the least impressive thing in AI right now, and quietly one of the most useful. It's a folder of plain text notes on your own machine, with one difference: a model can search across them on demand and surface the few that actually matter, instead of swallowing the whole thing. It fixes the most annoying problem AI has today, that it forgets everything between sessions and starts from zero every single time.
Everyone on AI is obsessed with building a second brain: every idea, meeting, and half-finished thought in one place a model can reason over. The exciting part isn't storage. It's that the model connects things you never connected yourself, and pulls up the right old note exactly when it matters. The personal version is basically a solved problem now. The open question is the team version. If one vault helps one person get more out of a model, connect the whole company's vaults and you should get a company that thinks better than any single person in it.
Meta built one for sixty thousand employees. Lyzr sells it as "Organizational General Intelligence." Strip the branding and it's the same move every time: pour everyone's knowledge into one tank, then ask the tank. Scraping everyone's Slack, Notion, meetings, and other tools into one vault gives you a searchable record of what the company already said; connecting people's separate vaults gives you what none of them could see alone.
Three of us connected our vaults and asked the simplest question first: where do we agree? Five patterns came back looking airtight. Each one appeared to have been reached by all three of us independently. Independent agreement across three separate minds is the strongest signal you can get that something is real.
So I ran the check I run on every project: an adversarial agent whose only job is to attack the result and tear it down. Inside an hour, it had the best finding on the floor.
The agreement was never independent. We work at the same company. What looked like three people reaching the same conclusion separately was mostly shared company canon leaking across the vaults. I didn't catch that. The agent I pointed at our own results did.
Agreement is the cheapest thing a team produces.
That agent existed because of an earlier faceplant. Back in April I ran a cruder version: one pass, no safeguards. It produced twenty-four confident "connections." I read every one by hand. About 80% were the kind a decent consultant could rebuild in thirty seconds from the same documents: true, and useless.
I built the adversary agent because that first run embarrassed me. The first thing it did at full strength was embarrass me again, faster. That's when the project actually started, because the agreement was never the valuable part. When people who work together agree, you've mostly found the habits they share.
The real value was in the opposite place: the gaps, where the three vaults didn't line up at all.
One example. I've been building AI agents longer than the other two. The newest member's vault held a complete cold-outreach playbook mine didn't: an entire way of working I would have sworn I already had covered. The most experienced person on the team learned the most from the newest one. That's the opposite of what people expect when they connect a team's knowledge. It turned out to be the whole point.
A merged brain makes a flatter company.
Separate vaults, connected through an abstract layer, are where real collective intelligence starts.
A run works like this. Each person exports a private-stripped abstract layer of their vault. Three agents compare every pair: one hunting agreement, one hunting conflict, one hunting what's unique. A triangulation pass keeps only what survives across the system. Then the adversarial agent tries to destroy what's left. Only the survivors make the report.
What survived was more useful than I expected.
A disagreement neither of us knew we had: I treat the system prompt as the durable part of an AI agent. A teammate treats the human-AI conversation as the durable part. We both assumed we were aligned on how to build agents. We weren't. If we had deployed on both assumptions, the agents would have fought each other in production. We caught it on a comparison screen instead.
Some methods did show up in all three vaults: defining things by what they are not, measuring everything, building a layer that learns from the measurements. The adversary let those stand but marked them cheap. That's just canon, the stuff anyone in our orbit absorbs without noticing.
The convergence that counted showed up somewhere stranger: only between people from different fields, never between the two who shared one. A way to decide whether an idea is worth building, broken into a handful of measurable checks. Two people reached it from opposite domains. The two in the same domain never did. A pattern that's everywhere is a habit. A pattern that only survives across a gap is a finding.
Roughly twenty places where one person had half of something and another had the other half. My agent engineering plus a teammate's work on AI personas is a complete agent. Neither of us would have built it alone.
Now hold that against what the rest of the field is building.
Pour everyone's knowledge into one brain and you get something that knows an enormous amount but can't tell what matters. Depth perception needs two eyes, slightly apart. Your brain reads the small difference between them and turns it into distance. Close one eye and the world goes flat.
A merged company brain is the one-eyed version.
A cyclops. Huge memory. Zero depth.
And the flatness is built in. You can't patch your way out of it. Merge the vaults and you lose, all at once, the things that made the group smart:
No outside view is left to catch a blind spot, because now everyone shares the same one.
You can't tell real agreement from copying, because nothing is independent anymore.
The two halves that used to complete each other get blended into one pile.
Pay enough money and you can get twelve sharp people to think like one average one.
So we built it the other way. The vaults stay separate, only the abstractions travel, and every person approves what leaves their machine first. What you connect is the difference between people.
That's the part one big brain throws away.
The question we started with is the one any team should ask before trusting something like this: does joining our vaults actually make us smarter together?
We tested it four ways:
Do we surface thinking none of us had alone?
Where do we genuinely agree?
Where do we complete each other?
Does the group get smarter over time?
The answer came back yes on all four. But that answer came with a catch. Every yes only held because the vaults stayed separate, and because agreement only counted after the adversary cut out the contamination.
So I want to be honest about what we haven't proven.
It's three people at one company. Close colleagues see roughly the same world, so the differences we found are real but narrow. The real test is connecting people from completely different fields and companies. We haven't run that yet.
The rules for who owns the shared layer, and who settles a fight over what gets added, are mostly unwritten.
We think connected vaults grow knowledge faster than any single brain can. We've watched the method improve itself across two runs. That's promising. It is not proof.
That honesty is the part the big-tank version can never copy. A company-brain sales deck will never include the slide where its own method caught it cheating. This piece does.
Every argument here went through the same adversarial agent as that first false convergence, told to break it rather than agree. The holes it found are still here, named.
That's the bet: not one shared brain, but a network of separate ones. Connected vaults.
I'd rather tell you what it can't do yet than sell you the flat thing that demos well.
So before you let anyone build your company a brain, or before you build one yourself, sit with a few questions.
Does it merge our knowledge into one place, or connect separate ones?
When two of us disagree, does it show me the disagreement, or average it into one safe answer?
Can it tell people who arrived somewhere independently from people who just copied each other?
Six months from now, is its answer sharper than the day we filled it, or staler?
Answer those honestly, about any company brain you're sold or tempted to build, and the merge starts to look like what it is: a confident way to go blind.
Don't build one big dumb brain. Connect all the smart ones you already have. That's the collective intelligence everyone's trying to buy.