Tell ChatGPT once that you're vegetarian, and a restaurant search months later can come back meat-free without another explanation. Tell it what you do for work and it stops asking. When memory is on, relevant details from past conversations follow you into new ones, so a fact said once can carry forward instead of being re-explained on every chat.
It comes on by default and works on the free plan. OpenAI's exact controls vary by plan, region, and platform, but the core skills are the same everywhere: seeing what memory holds, and fixing it when it holds something wrong.
The two halves of memory
Memory has two parts, and knowing which one produced a detail matters when you want to change it.
Saved memories are the facts you hand over directly. Say "remember that I'm a freelance designer" and ChatGPT stores that as a standalone item. Saved memories live separately from your chat history, a detail that matters later, when something needs deleting.
Reference chat history is the automatic half. With it on, ChatGPT can pull relevant details from past conversations without being asked: the project you described last week, the format you decided on earlier. There's no separate storage limit on this part. ChatGPT decides per response what's relevant, so it won't use everything, and a detail won't show up in every answer. The same mechanism personalizes web searches: "restaurants nearby" respects the dietary restriction you mentioned in March.
See what it remembers
Checking memory is the fastest way to catch a wrong detail before it shapes another answer. Settings, then Personalization, then Memory, is where the check happens. The memory summary there is a high-level view of the details ChatGPT may use, editable and deletable on the spot, and it can omit details. For a fuller picture, ask ChatGPT what it remembers about you.
When a response draws on your personal context, a Sources list can appear under the answer, pointing at the past chat, saved memory, or custom instruction that shaped it. Sources don't show every influence, but they help trace a response that assumed the wrong thing back to where the assumption came from.
Fix it or remove it
ChatGPT offers three ways to stop a detail from shaping future answers, and they do different things.
"Don't mention this again" tells ChatGPT to stop referencing a detail in future responses. The underlying chat or memory stays put; only the references drop.
Deleting a saved memory, or editing the memory summary, removes the item from future personalization. The asymmetry to remember: deleting a chat does not delete a saved memory made from it. To retire a detail fully, delete the memory and the chat it came from. OpenAI notes that logs of deleted saved memories can be kept up to 30 days for safety and debugging (Memory in ChatGPT).
Turning off Reference chat history disables the automatic half. Information remembered from past chats gets scheduled for deletion within 30 days, while saved memories stay, and so do the original chats.
And when a conversation shouldn't touch memory at all, start a temporary chat. Temporary chats don't create or update memories, and picking Unpersonalized before starting keeps the conversation from using your existing memories either. Saving one later converts it to a regular chat under your account's settings.
Where memory earns its keep
Memory pays off when a detail should follow you everywhere. Three uses tend to earn their keep fastest:
- Preferences that should apply everywhere. "I'm vegetarian." "No exclamation marks in work emails." "I use metric units." Said once, respected later.
- Standing context. What you do for work, the tools your team runs on, who you write for.
- Recurring project details. The client's style guide, the report format, the tone that took three drafts to find.
The payoff is cumulative: fewer setup paragraphs, fewer corrections, more room for the task itself.
Memory is also account-wide, so work details can surface in personal requests and the reverse. When context should stay scoped, put it in a Project instead: instructions saved inside a Project apply only there, so the client context stays with the client work.
One privacy note: on personal accounts with "Improve the model for everyone" switched on, OpenAI may use chats and remembered information to improve its models. The switch lives in Data controls. For conversations that shouldn't feed memory at all, a temporary chat with personalization off handles it at the source.
FAQ
What does ChatGPT remember about me?
Open Settings, then Personalization, then Memory: the memory summary shows a high-level view of the details ChatGPT may use. The summary can omit details, so asking ChatGPT what it remembers about you fills the gaps.
How do I turn off ChatGPT memory?
Settings, then Personalization, then Memory. Two separate controls live there: Reference saved memories (turning this off also turns off chat history reference) and Reference chat history, which you can manage separately when saved-memory reference stays on. Turning off chat history schedules information remembered from past chats for deletion within 30 days; saved memories remain until you delete them, and your chats stay in your history.
Does deleting a chat delete what ChatGPT remembered from it?
No. Saved memories are stored separately from chat history, so a memory made from a chat survives the chat's deletion. Delete the saved memory in Memory settings as well.
Is memory the same as Custom Instructions?
No. Custom instructions are direct guidance you write about how ChatGPT should respond. Memory collects relevant details that develop across conversations and other sources, like files in your Library or connected apps.
Does ChatGPT memory work on the free plan?
Yes. Memory is on by default for personal accounts, free plans included, though the exact controls vary by plan and region.
Want to go further?
- If memory keeps missing things inside one stream of work, What Is Context in AI? Why It Makes or Breaks Your Results covers what AI sees within a single conversation.
- If you'd rather not depend on memory at all, How to Build a Personal AI Workflow (Step by Step) turns standing context into saved instructions.