A way to give the whole company one shared memory it can actually trust — one that can always tell you how it knows what it knows, and never quietly makes something up.
What it is · why it matters · how it touches your work
01
The problem
A company’s truth is scattered everywhere — billing, the CRM, email, chat, documents, code, analytics. It’s in different shapes, some of it is sensitive, and it changes every single day. And what everyone actually wants is simple: to ask a question in plain English and get a straight answer. What’s the story with this account? What did we decide about that? What’s our position here?
The trouble is, the moment you point today’s AI at that pile, four things go wrong at once.
It can’t show its work
Ask where a fact came from and it can’t tell you. A real fact and a confident guess look exactly the same coming out of it.
It’s out of date
And it’s stalest right where it hurts most — a balance, a price, who the contact is now.
It leaks
It tells everyone the same thing — including things a particular person was never meant to see.
It gets more convincing, not safer
As the AI gets smarter, its mistakes get more believable, not less. The thing that should make it safer quietly makes it riskier.
So the real problem was never “how do we search all this.” Searching is easy. The problem is building a company memory that’s trustworthy on every one of those counts at once — and sturdy enough to run like plumbing, instead of a delicate machine you have to baby. Almost nothing out there clears that bar, because the usual approach isn’t built to.
02
Why the usual approach falls short
The obvious move is to pour everything into one big AI search index and let it answer. It demos beautifully and breaks on all four counts. It can’t show its sources — any citation it gives is made up after the fact and never checked. It’s either expensive to keep current or it’s stale. It decides what you’re allowed to see only after it has already read the sensitive material. And that search index quietly becomes the only copy of the memory — so when it drifts, there’s nothing to rebuild from.
You can’t patch your way out of those. You have to build it differently from the first day.
03
The idea — every fact carries how we know it
Here’s the whole thing in one move. Instead of storing “the truth,” the memory stores claims — and every claim comes with how we know it. Take a plain fact: Acme’s plan is $4,100 a month. How do we know that? It turns out there are only three ways anything ever gets into the memory:
most solid
We saw it happen
A $4,100 payment came through. That isn’t an opinion — it’s something that actually happened, and we recorded it. (The system calls this observed.)
↓
a step removed
We worked it out
It read the last few invoices and concluded the plan is $4,100. A fair read — but one step away from seeing it directly. (Inferred.)
↓
someone’s word
Someone told us
An account manager typed it in. Could be current, could be stale — it’s a person’s say-so, so it’s held the most lightly until something firmer backs it up. (Asserted.)
That one move — tagging every fact with how we know it — is what makes everything else possible:
It never paints over the past
When the price changes, we don’t erase the old $4,100 — we add a new fact that takes its place. The current answer is just the most recent one we’ve accepted, with a trail all the way back. Nothing true ever gets lost.
It remembers two different “whens”
Every fact knows when we learned it and when it was actually true. So “what did we know in March” and “what was true in March” are two different questions — and it can answer both.
It can’t fake being sure
Because it tracks how it knows each thing, it rates its own certainty honestly. A guess can never be dressed up as a hard fact — no matter how smart or smooth the AI gets.
The right person sees the right thing
Who put a fact in, and where it came from, also decides who’s allowed to see it. The gate is part of how it looks things up — so it can’t be talked into showing someone what they shouldn’t see.
And the same set of facts can be shown however you need it — a quick chat answer, a written report, a page on a site — with each one checked the right way for what it is. No version is its own separate truth; they’re all just views of the same record. Underneath, it’s a simple running list we only ever add to. The simplicity is on purpose.
04
Why it matters
None of this is a feature you switch on. It all falls out of that one idea — facts that carry how we know them.
It can always show its work
Every answer can tell you where each piece came from and how sure it is. The one thing the usual approach can only pretend to do, this one is built out of.
Privacy is built in
Not a rule bolted on afterward — a property of how the memory is stored and looked up. It can’t accidentally over-share.
It gets more trustworthy as AI improves
Most AI tools get riskier as they get smarter — the mistakes get more convincing. This one gets better, because it can never promote a guess to a fact, however clever it sounds.
It remembers how its own mind changed
When two sources disagree, or a fact gets updated, the whole history is still there — so it can tell you what changed instead of quietly picking one.
It’s sturdy, not delicate
The real memory is the running history. Everything people actually use is rebuilt from it — so if a piece broke, we’d rebuild it and lose nothing.
It’s worth more every day
Every new fact makes it more useful — and because it’s built right, that value doesn’t rot.
In one line: it turns the company’s scattered, ever-changing, sensitive truth into one memory the team can actually trust — one that can always say how it knows, shows each person only what they should see, and only gets better as the tools around it do.
05
Where it goes next
This isn’t a theory. We’ve already built this same kind of memory twice — once for a single practice’s truth, and once for a person’s whole world. The work ahead is mostly bringing what those proved to the company.
ALREADY WORKING
The hard part is proven
The same model — facts that carry how we know them, with a clean history — is already running. The company version is that, wider.
NEXT
Honest about getting stale
Teach each fact a “good until,” and have the answer say so when it’s old — and when it doesn’t know something, say “I don’t have that yet” and go find out.
THEN
How things connect
Teach it the relationships — which client leans on which vendor, who decided what — so it can reason across the company, not just look one thing up.
06
Want to build this into your project?
This pattern isn’t only for one team. Any project that needs to answer questions from a body of truth — and be trusted doing it — can use it. There’s a full build guide you can hand straight to your coding agent: the design as concrete steps and code, every promise written as a test that has to pass, what to reuse so you don’t reinvent it, and the research behind each choice.