What does it know about my company tomorrow morning that it knew tonight?
The promise that prompts it is the one every demonstration ends on: it learns your business. You nodded, because that is what you would want, and because the alternative sounds like a machine you must re-brief every morning. The honest position sits closer to the re-briefing, and knowing that decides whether you are buying a product or a filing habit.
Nothing survives the night on its own
A model has no memory between conversations. Each one begins blank, with no trace of the last, and nothing your staff said yesterday altered a single thing inside it. Everything that resembles memory is bookkeeping: text written down after one conversation and handed back at the start of the next.
In ArkOne’s reference design for an executive agent, the Cabinet, that bookkeeping is the Minutes. A separate pass reads the finished conversation and extracts the decisions: what was decided, on what reasoning, what remains open. It runs afterwards, so a slow minute-taker delays the record and never the answer. The next conversation opens by reading back the eight most recent decisions and every initiative still open, a setting on the Register. Eight rather than eighty, because everything read at the opening of a conversation is paid for at the opening of every conversation. The paper on the Minutes takes it further.
So the answer to your question is a short list. It knows the decisions somebody wrote down. It does not know the argument that produced them unless asked. And it knows nothing of the conversation your sales director had with it yesterday afternoon, unless that conversation ended in something the extraction pass recognised as a decision.
What the promise sounds like when it is pressed
Suppose the question goes to a shortlist of three, each having built something real, and each answering in a breath before the demonstration resumes.
| What a good answer sounds like | What you will be offered instead | What that hides |
|---|---|---|
| Here are the lines it will read tomorrow. A person can open them, search them and correct them | It has long-term memory and learns from every interaction | A feeling in place of a file. Ask what a person would open to read it, and whether anyone on the call has |
| Decisions are extracted; the reasoning stays in the audit record and is fetched on request | We keep the full conversation history | A transcript replayed nightly, billed nightly, with three useful lines buried in an argument |
| What it knows about each colleague sits apart from what it knows about the company | Everything is stored in your instance | Where the data sits, not what is in it. Both answers can be true and only one of them was asked for |
Exhibit 1. Illustrative. Three replies about memory, and the bill or the blank each of the weaker two conceals.
The second row is the expensive one, because it sounds the most generous. Keeping everything is a decision to pay for everything every morning, and to hand the model a haystack when it needed a line. Ask what that costs at the volume your firm would run, then ask what the same conversation bills with caching on and off.
The third row is a good answer to a question you did not ask. Data residency matters and belongs on your list. It says nothing about whether what is stored is a decision, a transcript or a summary somebody generated once and never checked.
The demonstration to insist on
The follow-up is not a question at all. Ask them to open the actual text the agent will read tomorrow morning, on the screen, now, and read one line of it aloud.
What you learn is not really about memory. A vendor who can do this built the thing as bookkeeping a person can audit, the same instinct that produces a readable audit record and a working approval step. A vendor who cannot will offer a summary view, a dashboard, or an assurance that it is stored securely, each a description of something you may not open.
Then ask the second half plainly: can I correct a line. A wrong decision written into the Minutes is read back at the opening of every later conversation until somebody fixes it, with the same confidence as a right one. Where a person may write a correcting line, the fault heals in a minute. Where the memory is closed, it compounds quietly for a quarter and then explains an outcome nobody could account for.
The word to listen for is learns. Underneath it there is only writing and reading back, and both of those are documents, holding your company’s decisions, that somebody at your company should be able to open.
component: answer-card
Asked plainly
Does an AI agent learn my business over time?
Not in the way the phrase suggests. The model itself is unchanged by your conversations; it begins each one with no trace of the last. What improves is the material handed to it: decisions written down after each conversation, documents added to its shelves, instructions corrected by a person. All of that is bookkeeping you should be able to inspect.
What does an AI agent remember between conversations?
Whatever a design chose to write down and chose to read back. In the reference design described here a separate pass extracts the decisions once a conversation has ended, and the next one opens by reading back the eight most recent decisions and every initiative still open. Older decisions remain searchable but are not read aloud each time.
Is keeping the whole conversation history the same as memory?
No, and it is usually worse. A transcript replayed at the start of every conversation is paid for at the start of every conversation, and it buries the two or three decisions that mattered in the argument that produced them. A short record of what was decided is cheaper to read and easier for a person to check and correct.