This is how to think about what AI remembers

This is how to think about what AI remembers.

People often choose one of two stories: the chat remembers everything forever, or it forgets everything when the window closes. Neither story is a safe general rule. “Remember” can describe several different layers, and products can implement those layers differently.

Separate four boundaries

• Current-chat context: earlier turns available while the conversation is being answered. This is why an old preference can influence a later request in the same thread.
• Optional product memory: a feature that may carry selected information across chats when enabled. Its controls and behavior belong to the specific service.
• Saved account history: conversations or activity visible in an account. Visibility in history does not by itself explain retention, model improvement, or deletion behavior.
• Deliberate handoff: the short brief you create and choose to bring into a new chat. This is the boundary you control most directly.

These layers can overlap, but they are not interchangeable. Starting a new chat can clear irrelevant conversational context without proving anything about saved history or a product memory setting.

Use a harmless context experiment

The boundary prompt:

This is a fictional context exercise. In this chat only, remember that the weekend plan is for one adult, has a $75 ceiling, and must avoid driving after dark.

First, restate the three constraints you received in this conversation. Then create a clean handoff brief containing only the goal, those constraints, and one open question. Label which details came from my current prompt. Do not claim anything about account memory, saved history, or what a new chat would know.

Use fictional or harmless details only. The explicit source labels are the part that matters.

Do not test privacy with private information

If you want to understand a service’s memory, use a harmless invented preference, then inspect the product’s current official controls. Do not enter a secret to see whether it comes back. A successful recall experiment would not tell you every place the data may be stored, and a failed recall would not prove deletion.

For important work, keep your own protected source of truth and carry forward a short safe brief. Do not depend on a chat to remember a commitment, and do not depend on a new chat to forget something you should not have entered.

The practical question is not “Does AI remember?” It is “Which layer am I using, what does this product currently say it does, and what information did I choose to place there?”

• Share only: Fictional examples and short harmless context; do not probe memory by entering secrets, identity details, confidential work, or sensitive records.
• Verify: Actual memory, history, retention, deletion, training, and privacy behavior in the current service’s official settings, policy, and documentation.

This is how to use AI.