Your Agent Is Keeping a File on You. Nobody Agreed Who Owns It.

OpenAI quietly changed what your assistant remembers. Its updated memory system, internally called Dreaming, stopped saving scattered bullet points and started building coherent narrative profiles, sorted into categories like work, hobbies, and travel. The success rate for keeping that profile current jumped from 52.2 percent a year ago to 75 percent, according to The Decoder.

Read that as a product update and it sounds helpful. Read it as commerce infrastructure and it is something else. The assistant that is about to buy things for you now holds a structured file on who you are. That file is the buying context. And no one has settled who owns it.

The agent does not shop with your preferences. It shops with a model of your preferences, held by a company that is not you.

What actually changed

The difference between scattered notes and a dossier is the difference between a shoebox of receipts and a credit file. One is data. The other is a profile that can be reasoned over.

Scattered memory said "user likes hiking." A dossier says you are a lapsed premium cardholder who travels for work, buys outdoor gear in spring, and abandoned two carts over shipping cost. The first is trivia. The second is a sales brief. When ChatGPT crossed from one to the other, it crossed from remembering you to representing you.

This is the memory layer of the agentic stack maturing in public. We have written about how agent infrastructure is shifting toward persistent memory and state. The dossier is what that memory looks like once it is good enough to act on.

Why this is a commerce problem, not a privacy footnote

The instinct is to file this under privacy and move on. That undersells it. The dossier is becoming the thing that decides what you are shown and what gets bought.

When you ask an agent to find a running shoe, it does not query a neutral catalog. It queries one through the lens of your profile, which means whoever holds the profile shapes the result. The merchant sees a buyer. The agent sees a brief. The profile sits in between and tilts the outcome. That is commercial power, and it has quietly relocated from the merchant's CRM to the model provider's memory store.

The data industry already understands this, which is why the data layer is being rebuilt for agentic commerce. We covered the Publicis and LiveRamp move to assemble a data layer for agent-driven shopping. The dossier is the consumer side of the same contest: the most useful purchasing data in the world is being assembled inside the assistant, one conversation at a time, with no clear owner.

Name the gap: you cannot move it or fully consent to it

Two things are missing, and both are load-bearing.

The first is portability. Your dossier is locked to the provider that built it. Switch from one assistant to another and the model of you does not come with you. You start over. For a profile that took a year of conversations to assemble, that is not a minor inconvenience. It is a lock-in mechanism dressed as a feature. The richer the dossier, the harder it is to leave, which is exactly why the provider is investing in it.

The second is consent. There is no standard for what an agent may do with the dossier when it acts in a commercial setting. The memory was built from a chat about your weekend. It is now informing a purchase. Nobody drew the line between those two uses, and no receipt records which version of you authorized the buy.

Put this against the MM Trust Layer Model and it lands on the authorization layer, the one we keep flagging as unfinished. We argued in our piece on the agentic payments identity crisis that the system cannot yet say who an agent is acting for, with what standing. The dossier sharpens that question. The agent is acting for a profile of you. When the purchase is wrong, which version of you is on the hook?

The counter-model is already shipping

This is not settled, and the alternative is not theoretical. OpenJarvis, an open-source framework out of Stanford, runs the whole personal-agent stack on-device, including memory. The bet is explicit: keep the dossier local, under the user's control, rather than in a provider's cloud.

That is the fork. One path keeps your profile in the assistant and makes it the provider's asset. The other keeps it on your device and makes it yours. Both are being built right now, and the commercial incentives point hard at the first. Local-first memory is better for the user and worse for the business model, which tells you which one gets the marketing budget.

As Simon Taylor has argued in Fintech Brainfood, the agentic commerce stack is turning out vertical, with each layer owned by a different player rather than shared as an open rail. The dossier is the clearest case yet. Whoever owns the memory owns the buyer, and right now that owner is the model provider by default, because no one wrote a rule saying otherwise.

What to watch

Watch for a portability standard, or the conspicuous absence of one. If the providers ship rich memory and no export, the lock-in is the point. Watch the first regulator to treat an agent dossier as regulated personal data with a right to move it. And watch whether any merchant or network tries to build the neutral profile, the one that belongs to the shopper rather than the assistant, because that is the opening.

The assistant got better at remembering you this week. The harder question is the one nobody answered. The file exists. It is about to spend your money. Who owns it?

The agent will shop with a model of you that you have never seen. Would you sign off on a purchase made by a version of you that you cannot read?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.