What we keep out of our AI's memory, and why
This week we copied our company memory so a second AI provider could read it. Deciding what to leave out was the hard part.
3 min read Provecta Group
On Wednesday we built a copy of our company memory for a second AI provider to read. The reason was practical. Our rules, facts and past corrections had grown up around one AI tool, and we wanted the same memory available whichever model we use. Relying on one vendor for something this central is a risk of its own.
Making the copy was the easy part. Deciding what must not be in it was harder, and it is the more useful part to share.
What never goes in
Credentials of any kind. Passwords, API keys, access tokens, bank logins. The export step strips every line that looks like a credential before anything leaves the machine, and after the first run a sweep of the copy found none left. We do not rely on the sweep alone. Credentials are not supposed to be in the memory in the first place. They live in a separate store, and the memory only says where to ask.
Material that belongs to a different company. We work across more than one company, and one of them keeps an entirely separate memory of its own. It is not in the copy at all. A second tool we built this week, which syncs plans and tasks from the memory into our back office, reads only the one company's folder it is meant for.
Anything we would not show a careful new hire. That is the working test. If a file would be wrong to hand to a capable person on their first day, it is wrong to hand to an AI provider whose data handling we know only from its terms.
What everything in it carries
A company. Every rule and record belongs to one company, and the rules start by naming which.
A reason and a date. Each rule says why it exists and when. That makes stale or wrong entries easier to find and easier to remove.
A history. The memory is kept in a version-controlled store the company owns, so every change shows what changed and when. If a bad fact gets in, we can see how, and take it out.
Why this matters more than it looks
IBM's 2025 Cost of a Data Breach report found shadow AI involved in 20% of the breaches it studied, adding about USD 670,000 to the average cost, and 63% of organisations had no AI governance policy. Most of that risk comes from data going into AI tools with no structure at all.
A company memory is the opposite of no structure, which is why it helps. It is also a single, rich store of what the firm knows, which is why it needs care. Once that memory is read by more than one AI provider, every decision about what is in it is made several times over. A password left in a file used to be one exposure. In a memory loaded by several tools, it is several.
The legal side
If your memory holds personal data about clients, staff or prospects, data protection law applies to it like any other record. Under Kenya's Data Protection Act, 2019, and GDPR in Europe, you need to know what personal data you hold, why, who processes it and for how long. Sending a memory that contains personal data to an AI provider is sharing it with a processor. Check that provider's data terms and where it stores data before you do.
The simplest way to stay on the right side of this is to keep personal data out of the memory unless a task truly needs it. Rules, methods, prices and lessons rarely need a person's name.
What this approach will not do
Redaction catches patterns. It can miss a secret written in an unusual format, which is why the real rule is not to put secrets in at all. Keeping a company out of the copy does nothing about a person pasting that company's material into a chat by hand. And a careful memory does not make an AI provider's systems secure. It only limits what they hold.
We also have not had an outside party audit this. It is our own practice, checked by our own sweep.
A five-minute check
Search whatever you use as shared AI instructions, whether a custom assistant, a project file or a pinned document, for the words "password", "key", "token" and "PIN". Then search for the names of three clients. Anything you find is the first thing to move out.
Start with the AI readiness check for a wider view of how your AI use handles data. Book 20 minutes at pgco.world/book.