Your Company's Memory

If you have just installed Obsidian, Claude and a folder full of company documents, the setup guide got you to the starting line. This is about everything after it — the handful of decisions that determine whether the thing is still useful in a year, or a folder nobody opens.

None of it is complicated. Most of it is unglamorous. All of it is the part people skip.

1. Start here: it does not know your company

This is the one idea everything else hangs on, so it is worth being precise about.

The model does not hold your business in its head. It never learned your company and it remembers nothing between questions. Every single time you ask it something, the same thing happens: it searches your documents, pulls out a handful — call it four to eight — and writes an answer using only those. Then it forgets.

2. So “hallucination” is almost always an input problem

When the right page does not come back from that search, the model does not stop and tell you. It answers from general knowledge instead, in exactly the same confident voice it uses when it is right. You cannot tell the two apart by reading them.

In real companies it comes down to three things:

• Contradiction. Two price sheets, neither one dated. It picks one. Nobody ever finds out which.
• Missing. It was never written down anywhere — it lives in your foreman's head. So the model fills the gap itself.
• Buried. A ninety-page contract stored as one file. It retrieves the wrong section and answers from that.

The model is rarely the problem. The filing cabinet is.

3. Five things that make it trustworthy

1. One source of truth per topic. Find the duplicates and delete them. Ruthlessly. This unglamorous step is most of the fix.
2. A date, an owner and a status on every document. Who owns it, when it was last true, whether it is current or superseded.
3. Archive the superseded. Move it somewhere the system does not read. Old documents left sitting in place are the single biggest source of confidently wrong answers.
4. Break big documents into pieces. Topic-sized. A ninety-page PDF retrieves badly no matter how good the model is.
5. Demand a citation on every answer. No source shown, no use. If it cannot tell you where it got something, you cannot act on it.

4. Five things that will burn you

1. Do not dump everything in. More documents does not mean better answers. Noise actively makes retrieval worse.
2. Do not put podcasts and books in a company brain. That is personal learning. It dilutes the pool and pollutes answers about your business.
3. Do not let it do arithmetic. Language models are confidently bad at math. Numbers come from your system of record.
4. Do not let everyone write to it. More on this next — it is the mistake with the longest tail.
5. Do not build it once and walk away. An unmaintained system does not go quiet. It goes confidently out of date.

5. Everyone reads. Almost nobody writes.

Reading from a company brain is easy and mostly harmless. Writing to it is where companies get hurt, and it is the question almost nobody asks before they start.

Separate the two, and split what goes in by how much authority it carries:

If anyone can write to it, in six months it is full of one person's opinions — and it looks official.

6. A brain built once is worse than none

Your SOPs changed. You signed new contracts. You ran forty more jobs. The vault did not.

It does not announce this. It keeps answering, in the same confident voice, using last year's truth. That is worse than having nothing, because nothing does not mislead anybody.

Three habits prevent it, and together they cost about an hour a quarter:

• A named owner. Every document belongs to a person, not to “the company.”
• Quarterly confirmation. Each owner confirms theirs is still true. Fifteen minutes, four times a year.
• A stale flag. Anything unconfirmed gets marked, and the system says so when it uses it.

7. If you work in finance, four rules

1. Never let it calculate. Numbers come out of the ledger. The system retrieves and cites — it does not do arithmetic.
2. Separate “what is our policy” from “what is the number.” It is excellent at the first and genuinely dangerous at the second.
3. Nothing client-facing goes out unverified. A human opens the cited source and confirms it. Every time, no exceptions for speed.
4. Client data does not live in a personal cloud drive. Ask where the vault actually sits, who else can reach it, and what happens when a laptop walks off. Ask before you load anything, not after.

The tool has no idea when it is wrong. That is entirely your job.

8. Your newest hire is your best auditor

Do not hand a new person a search box and wish them luck. Give them ten questions to ask it in their first week — the ten things everybody at the company had to learn the hard way.

Then watch for every time they give up and go ask a human instead. Write those down. That list is precisely what your system is missing, ranked by what a real person actually needed on a real day.

Do that and onboarding stops being a cost. It becomes a free audit you run every time you hire.

9. The boring infrastructure nobody builds

If you get past the basics, these four are what separate a system people trust from one they quietly stop opening:

10. Start Monday. It is smaller than you think.

You do not need any of the above to begin. You need one topic and about two hours.

1. Pick one topic. Pricing, or your ten most-used SOPs. Not everything — everything is why people quit in week two.
2. Find every version. Keep one. Date it, name an owner, move the rest to an archive folder the system never reads.
3. Ask it ten questions you already know the answers to. That is your baseline. Re-run it monthly and watch whether it drifts.

If the answers to those ten are good, widen the topic. If they are not, the fix is almost never the model — it is something in this article.

The tools are free. The work is what makes them worth anything.