ChatGPT Memory — how it works and how to control it

Where the controls are

Open Settings → Personalization → Memory. Depending on your account you'll see some of: Memory, Reference saved memories, Reference chat history, Memory summary, Manage or Saved memories. Controls vary by plan, region, platform and workspace, so a missing toggle usually means the feature isn't on your account yet.

The two mechanisms

• Saved memories: details you ask ChatGPT to remember, or that it saves as useful context. They are stored separately from chat history. You can ask it to forget one or delete it in settings. Depending on plan, you may also get search, sorting, automatic prioritisation and version history.
• Reference chat history: lets ChatGPT use relevant information from past conversations. That derived context changes as ChatGPT updates what's useful. Turning it off schedules the remembered information for deletion within 30 days, but your original chats stay unless you delete them. Turning off Reference saved memories also turns off chat-history reference.

Some accounts can switch between improved memory (a continually updated summary) and the legacy saved memories list. Switching doesn't delete your history.

See what it knows

• Memory summary: a high-level view you can correct inline, mark Don't mention this again, or wipe with Delete and turn off memory (which does not delete past chats). OpenAI notes the summary may omit details.
• Sources: when personal context shaped a reply, Sources may appear under it: a past chat, saved memory, custom instruction, file or connected email. You can open, correct or delete from there. Sources aren't included in shared links.
• You can also just ask: "What do you remember about me?"

How to actually remove something

Information can live in several places, so delete it from each one that applies:

1. the memory summary or the saved-memories list;
2. the chat where you first shared it, including archived chats;
3. files in Library;
4. connected apps. Disconnect the app to stop future access.

Don't mention this again only changes personalisation; it doesn't delete the source. Deletion can take time to propagate, and OpenAI may keep logs of deleted saved memories for up to 30 days for safety and debugging.

Temporary chats

Before starting a temporary chat you choose whether it uses existing memories, custom instructions and plugins. Choose Unpersonalized to leave them out. You can't change that after the chat starts. Temporary chats never create or update memories. If you save one, it becomes a regular chat and follows your normal Memory settings.

Privacy and training

• Personal accounts: chats and remembered information may be used to improve models while Improve the model for everyone is on. Turn it off in Settings → Data controls.
• Business, Enterprise, Edu and ChatGPT for Healthcare: not used for training by default. Workspace owners can manage memory settings and role permissions.
• Regulated workspaces (Healthcare, Enterprise with Regulated Workspace) have improved memory off by default. It isn't covered by the BAA, so don't enter PHI. Project-only memory can confine context to a single project.
• Memory may use relevant details to refine web searches (for example, your location for a restaurant search).

Memory vs custom instructions

Custom instructions are direct guidance you write, applied to every chat. Free and Go allow up to 1,500 characters and paid plans 5,000 (Custom Instructions). Memory is what ChatGPT picks up across conversations and other sources. Use instructions for rules and style, and Memory for context.

How other agents handle memory