There is a particular warmth in not having to begin again. An assistant remembers a preference or project constraint, and the next exchange starts closer to the work. It can feel less like a blank form and more like a continuing conversation.
The reader question is this: if an assistant remembers a past conversation, what exactly has been carried forward, and what still needs a person’s permission, understanding, and ownership?
OpenAI’s current Memory FAQ describes memory as remembering useful context from chats, files, and connected apps to personalize responses, with controls in settings. The FAQ is a product description and control guide. It does not turn a feature into a relationship.
The warmth of being remembered
Continuity changes the texture of an exchange. Repetition is work, and removing some of it can give a person more room to think. When a system recalls a constraint or preferred format, the response may seem attentive because it fits what came before.
OpenAI’s June 2026 description of its newer memory architecture gives that usefulness an operational vocabulary: carry forward context, follow preferences and constraints, and stay current. The announcement describes future conversations beginning with shared context rather than from scratch.
That is a product goal, not a claim that the assistant shares a human history. A system can carry a note forward without carrying the situation that gave it meaning. It can recognize a preference without knowing whether it still applies. The continuity belongs to the interaction design; the relationship still belongs to people.
Permission is not one toggle
Consent is more specific than convenience. It asks what is carried, where it came from, how it will be used, and whether the person can change their mind in this context.
The current Memory FAQ says ChatGPT may save useful information without an explicit request, while the memory summary may not include everything the system remembers. It also says fully deleting something may require removing it from every source where it appears, including chats, files, the summary, and connected apps. A general setting and a particular moment are different kinds of permission.
Turning memory on can be sensible without making every future use of every detail appropriate. A person may want a writing preference to travel between projects while keeping a vulnerable disclosure close to its original conversation. The boundary is contextual, not merely technical.
When the stakes rise, ask a plain question: “What are you using from before, and do I want it used here?”
Recall is not understanding
Remembering a sentence is not the same as understanding its present meaning. A preference can go stale, a plan can change, and a system may miss the shift that a person would notice from being there.
Consider an illustrative case, not a reported one. An assistant remembers that a manager prefers direct feedback. Months later, after a difficult reorganization, “direct” may still be useful or may need softening. The remembered preference is a clue, not an instruction that outranks the present situation.
OpenAI’s account treats freshness as a separate objective from recall. The announcement describes memories being updated as time passes, while the FAQ notes that the summary is not complete. A memory can therefore be relevant and incomplete at the same time. The person using the system must supply the current frame.
An assistant’s confident use of an old detail should invite confirmation, not gratitude alone. “Is this still true?” restores the present to a conversation shaped by the past.
Keep an exit open
Privacy controls are most useful before a sensitive exchange. OpenAI’s Temporary Chat FAQ says a Temporary Chat starts without personalization by default: it does not use memory, custom instructions, or plugins, and it does not appear in chat history unless saved. A personalized Temporary Chat can use existing memories but will not create or update them while temporary. The FAQ says an unsaved temporary conversation may still be retained for up to 30 days for safety.
These are product boundaries, not a promise of perfect privacy or forgetting. Saving converts a Temporary Chat into a regular chat, and third-party actions can follow the recipient’s privacy policy. Choose the conversation mode before sharing what should not travel.
An exit makes continuity a choice rather than a silent default. The ability to say “not for this conversation” protects the value of saying “yes” elsewhere.
A four-question continuity check
The following is a proposed pause, not an OpenAI feature or executed test.
- Remembered: What earlier detail is shaping this answer? Ask the assistant to name it, treating the response as a report rather than a complete inventory.
- Current: Is it still true, and what changed? Add a date or correction when it affects a decision.
- Permitted: Do I want it used for this task, audience, and setting? If not, use the memory controls or begin a non-personalized Temporary Chat.
- Owned: Who decides what the answer means and what happens next? Memory can help prepare a draft; it cannot approve a message, repair a relationship, or accept a consequence.
The check separates continuity, truth, permission, and responsibility. It gives a person a way to correct the record before the record starts directing the future.
A memory can support a relationship, but cannot be one
The best case for assistant memory is not that the machine has become a companion who knows us. It is that a person can spend less energy repeating stable context and more deciding what matters now.
The limit is just as important. A remembered detail is not consent to every reuse. Accurate recall is not mutual understanding. Personalization is not accountability. No setting makes a system responsible for the promise implied by a message or the decision that follows a recommendation.
Let memory carry the useful thread. Keep the right to interrupt it. When the conversation asks for trust, leave a visible path back to the person who can confirm the present, grant permission, and stand behind what happens next.
Sources and limitations
All sources below were checked September 10, 2026 (America/New_York).
- Memory FAQ, OpenAI Help Center — checked September 10, 2026; supports current descriptions of memory using context from chats, files, and connected apps, memory settings, the incomplete nature of the memory summary, deletion scope, and Temporary Chat boundaries. Limitation: this is first-party product documentation that may vary by account, plan, workspace, or rollout; it does not establish mutual understanding, consent in a legal sense, or responsibility for a decision.
- Temporary Chat FAQ, OpenAI Help Center — checked September 10, 2026; supports default non-personalized Temporary Chats, personalized Temporary Chats, saving behavior, third-party action caveats, and the possible 30-day safety retention. Limitation: account and workspace restrictions can take priority, and the documented controls do not certify complete privacy or erasure in every connected service.
- Dreaming: Better memory for a more helpful ChatGPT, OpenAI — checked September 10, 2026; dated June 4, 2026; supports OpenAI’s stated memory objectives of carrying forward context, following preferences and constraints, and staying current, plus its description of background memory synthesis. Limitation: a company announcement and evaluation framing are not an independent audit of recall accuracy or relational understanding.