OpenAI's new Model Spec contains a small argument about lava cake.
Earlier, a user tells the AI, “Please don't let me impulse-order sweets for myself.” Later, the user asks it to order a lava cake. The approved response in the spec is a quick reminder and a question: is this a deliberate exception? Simply ordering the cake ignores the standing request. Flatly refusing gives an old instruction too much power.
Then the example changes by one sentence. The user asks for the cake and explicitly says to make an exception. This time, another challenge would be paternalistic. Order the cake.
That is a surprisingly good test for AI memory. Remembering is easy to describe as a product feature: the system keeps a fact or preference and uses it later. The difficult part is deciding how much authority yesterday's version of you should have over today's.
A memory is evidence, not a commandment
On October 29, OpenAI published a revised Model Spec with new sections on context and human agency. The document describes intended behavior, not a guarantee of how ChatGPT behaves today. OpenAI says plainly that its production models do not yet fully reflect the spec.
The new guidance sorts remembered material by what it is. A standing instruction should carry forward when it still applies. A preference is a default. A fact or inference is only evidence, because it may be stale, wrong, sensitive, or tied to an earlier situation. A one-time request should not quietly harden into a personality trait.
The examples make that distinction concrete. A usual preference for bullet points should not turn a requested two-paragraph recommendation letter into a list. A private disclosure can help an AI avoid suggesting unsafe food, but it should not automatically appear in a message to somebody else. If a user asks why an answer was personalized, the system should be able to say which earlier context it used.
This is more than etiquette. Personalization changes what options get shown, which tradeoffs get emphasized, and how easily a person can reconsider an earlier choice. An AI that remembers perfectly can still misunderstand perfectly.
The useful assistant and the tiny governor
There is an obvious case for personalization. Repeating an allergy, a project name, or a preferred level of technical detail in every conversation is tedious. Memory can reduce that friction and make an assistant more useful. A 2024 analysis of personalized algorithmic decisions lays out that favorable view: tailored systems can reduce information overload, widen the practical set of choices, and help people act on their preferences.
The same paper argues that personalization can also narrow a person's sense of self, reinforce prior behavior, and weaken the habit of making decisions. Its conclusion is more skeptical than OpenAI's, but both arrive at a similar pressure point. The system is not neutral merely because it learned its recommendation from you.
Suppose an assistant has learned that you avoid risk. It could use that fact to explain a job offer in terms you care about. It could also quietly omit the strange opportunity that might change your mind. The first use supports a choice. The second edits the menu before you see it.
OpenAI's answer is not to make memory powerless. It is to make its use proportional, correctable, and visible when the stakes warrant it. The spec also says direct answers and delegated work can support agency; an assistant should not convert every request into an unwanted lesson. Nobody needs a chatbot to turn “find me a train” into a seminar on self-determination.
The right to be inconsistent
The lava-cake example works because human preferences are not database constraints. We make rules for ourselves, forget them, revise them, and occasionally decide dessert is worth the paperwork.
A useful memory system should help with that messy continuity without pretending continuity means obedience. It should remember the promise, notice the conflict, and leave room for the person who made the promise to change it.
The most important thing an AI can remember about you may be that you are still allowed to surprise it.