Where does what it remembers live, and can I read it?
Two years from now you may want to replace this vendor, or your general counsel may need to know what a machine was told about a client before it wrote to them, or a departing employee may ask what is held about her. All three arrive on a Tuesday with no warning, and all three turn on one fact: where the memory sits, and whether anyone at your company may open it.
A store, a shape, and a key
What an agent remembers is a document. In ArkOne’s reference design for an executive agent, the Cabinet, it is called the Minutes: decisions written down after each conversation and read back at the start of the next, described in the paper on the Minutes. Three properties decide whether that document is yours.
The store: which database holds it, and whose account pays for that database. The shape: whether a line reads as a sentence a person understands, or as an encoding only the vendor’s software interprets. The key: whether somebody at your company can open it without asking, search it, and add a correcting line.
Underneath sits the audit record, and it follows a different rule. It is append only, a setting on the Register: an entry is never edited, a correction is a new entry, and the original stands. That is what makes it evidence rather than notes. The Minutes are the short readable layer your staff work with; the record beneath holds what the minute-taker saw.
Custody, in three replies
Suppose the question is put to a shortlist in one week, and each reply is written down as it was said.
| What a good answer sounds like | What you will be offered instead | What that hides |
|---|---|---|
| It writes to a database in your account. Here is a line, in English, and here is a person editing it | Your data is yours, and it is fully encrypted at rest | Custody described as security. Encryption protects a store from strangers and says nothing about whether you hold the key to open it |
| You can export every line, in a readable format, without asking us | There is an export function in the settings | An export in a format nobody else reads is a copy you cannot use. Ask what the file looks like, and whether a person could read it |
| What is held about one named colleague can be removed while you watch | Deletion requests go through our support team | A process, not a control. Ask how long it takes and who approves it, then ask what you would tell an employee who asked today |
Exhibit 1. Illustrative. Three replies about where memory sits, and what each of the weaker two makes you accept.
The first row is the one that trips careful buyers, because encryption is a genuine answer to a genuine risk and your security lead will be satisfied by it. It answers who cannot read your memory. You asked whether you can.
The third row will matter sooner than you expect. What an agent holds about individual colleagues is what a person is most likely to ask about, and a design keeping per-person notes apart from company decisions can remove one without disturbing the other. Where the two are mixed, deleting what is held about a leaver means picking apart reasoning behind decisions still in force.
Two things to make them do, in the room
The follow-up is a demonstration in two parts. Show me the exact text the agent will read tomorrow morning, and then delete what it holds about one named colleague while I watch.
The first part tells you whether the memory is readable. The second tells you whether it is yours, because deletion is the one operation nobody builds unless they intended the customer to hold the controls. A vendor who does both in the meeting has answered several of your other questions at the same time, including the one about what it knows tomorrow morning.
Where either part is refused, ask what would have to be true for it to be possible, and listen for whether the obstacle is technical or commercial. A technical obstacle is a roadmap item you can put a date against. A commercial one is a switching cost being defended, and it is better to price it now, in the negotiation, than to discover it in the quarter when you have decided to leave.
Then put one line in the contract while they are still selling to you: a readable export of everything held about your company, on request, in a format your next supplier can read. That sentence is cheap to agree before signature and is the whole price of the exit afterwards.
component: answer-card
Asked plainly
Where does an AI agent store what it remembers?
In a database somewhere, which is the whole of the question: whose database, in what form, and who may open it. In the reference design described here the record of decisions sits in a store the company owns, as plain lines a person can read, search and correct, with the fuller audit trail beneath it.
Can I read and correct what an AI agent knows about my company?
You should be able to. A memory a person can open is one a person can check, and a wrong line otherwise gets read back with full confidence at the start of every conversation until somebody notices the pattern. In a design with an append-only record, a correction is a new line and the original stays visible.
What happens to what an AI agent knows if I change vendors?
That depends on where it was stored and in what format. Decisions kept as readable lines in a store you control leave with you. Memory held inside a vendor's own system, in a shape only their software reads, is a switching cost you agreed to at signature without it appearing anywhere in the price.