Saved memories are a list of specific facts you asked ChatGPT to keep, and you can open that list and edit it. Reference chat history is ChatGPT quietly reading your past conversations on its own, and it is the one responsible for ChatGPT knowing things you never asked it to remember. They are separate features, controlled separately, and they behave differently when you delete a chat.
First, a dating note: since 4 June 2026 OpenAI has called the saved memories setup the legacy saved memories system, and its Memory FAQ now files it under a heading that reads Legacy memory system. The current default is a single Memory control with an editable memory summary. Plenty of accounts still show the old pair of switches, so both layouts are live right now, and this post covers what each toggle did, what replaced them, and how to tell which system your account is on.
The difference in one line each
Saved memories are things ChatGPT wrote down. Reference chat history is things ChatGPT noticed. That distinction explains almost every behavior difference between the two.
- Saved memories: details you have explicitly asked ChatGPT to remember, such as your name, your dietary preferences, or how you want answers formatted. OpenAI describes them as working like custom instructions, except the model updates them automatically instead of making you manage every line by hand.
- Reference chat history: ChatGPT pulling useful information out of past conversations to make future ones more relevant. OpenAI's example is that if you once mentioned you like Thai food, it may factor that in the next time you ask what to have for lunch.
- Saved memories stay until you delete them. Details drawn from past chats can shift over time as ChatGPT updates what it considers worth carrying forward.
- ChatGPT does not retain every detail from past chats. OpenAI's own guidance is to use saved memories for anything you want it to always keep in mind.
If you want something remembered reliably, say the word remember. That routes it into saved memories, which persist until deleted. Anything you merely mention in passing lives in the reference chat history layer, where it can fade, drift, or get reweighted without warning.
What saved memories actually stores
Saved memories is a discrete, inspectable list. You add to it by telling ChatGPT something like remember that I am vegetarian when you recommend a recipe. ChatGPT may also save details on its own when you share something that looks useful later, which is where most people's surprise entries come from. Unless you delete them, saved memories are part of the context for every future response.
The storage detail that matters most: OpenAI describes the saved memories notepad as stored separately from your chat history. That one architectural fact is why deleting a conversation does not remove what ChatGPT learned in it. The memory left the chat the moment it was written.
ChatGPT can also manage that list for itself, combining entries, updating them, or dropping them when asked. Later updates to the older system added automatic prioritisation, keeping recent and frequently discussed details in the foreground so accounts stop hitting a memory full state. You can sort and search the list, and switch automatic management off.
What reference chat history actually stores
Reference chat history does not store a list at all. When it is on, ChatGPT looks back across your previous conversations and pulls forward whatever seems relevant to what you are asking now. OpenAI compares the effect to custom instructions: relevant information from past conversations may be added into new ones. The key word is may. Nothing here is guaranteed, and nothing here is a fixed record you can scroll through.
Behind it sits a background process OpenAI calls dreaming, introduced in April 2025, which curates memory automatically by referencing chat history rather than waiting for you to say remember. OpenAI's framing is that saved memories were written only during a conversation and only on strong cues, which made them go stale, while the background process learns across many conversations and keeps the picture fresher.
The two features fail in opposite directions. Saved memories are precise but go out of date, because a line written eighteen months ago carries the weight it had on day one. Reference chat history stays current but is imprecise, and you cannot audit a synthesis line by line.
| Behavior | Saved memories | Reference chat history |
|---|---|---|
| What it is | A list of specific details, mostly things you asked it to keep | ChatGPT drawing relevant context out of past conversations |
| Can you view every item | Yes, it is a browsable list in Settings | No, there is no complete item-by-item view |
| How you add to it | Say remember, or let ChatGPT save it automatically | You do not add to it, it reads what you already wrote |
| Does it change on its own | Kept until you delete it | Details can change over time as ChatGPT reweights them |
| Reliability | Always considered in future responses unless deleted | Not every detail is retained, recall is best effort |
| Survives deleting the chat | Yes, it is stored separately from chat history | Not in the same way, deleting the source chat takes it out of what can be drawn on |
| Causes the surprise effect | Sometimes, when ChatGPT saves without being asked | Most of the time, this is the usual culprit |
| How to stop it | Delete individual memories or turn memory off | Turn the setting off, then delete the underlying chats |
Which one makes ChatGPT know something you never told it
Reference chat history, in most cases. When ChatGPT mentions your job, your city, or a project you have not brought up in that conversation, the usual explanation is that it read an old chat, not that a memory was saved. You can check this yourself: if the detail does not appear in your saved memories list, it came from the history layer.
There is a way to confirm it rather than guess. Tap the book icon below a response to see the sources that personalised it, which can include custom instructions, past chats, files and memories. Tapping a memory there opens an explanation of why it was used. OpenAI notes this view may not show every factor, so treat it as strong evidence, not a complete audit trail.
One more source of the surprise: saved memories can show up inside old conversations. If ChatGPT knows you live in San Francisco, a past answer might read since you live in San Francisco, even for a chat where you never said it. Deleting the memory will not scrub those sentences from old threads. Only deleting the chat itself does that.
How to turn each one on or off, independently
If your account still shows the two-toggle layout, open your profile icon, then Settings, then Personalization. Both switches live there and can be set independently. OpenAI's own wording is that you can choose to turn on saved memories, both, or neither, and change it any time. Release notes give the fuller path for the history switch as Settings, Personalization, Memory, Reference chat history.
If your account has the updated system, the controls consolidate to Settings, Personalization, Memory, and you get a memory summary page instead of two switches. From the three-dot menu on that page there is a Delete and turn off memory option, which clears the memories shown and turns memory off without touching your past chats. OpenAI notes this control reached the web first and was rolling out to mobile after.
- Your phone and your laptop can disagree. The updated memory experience requires a current app version, so an outdated app can still be showing the old layout. Update it before concluding a setting has vanished.
- To get back to the legacy saved memories system on an updated account, go to Settings, Memory, and use the saved memories link under Memory summary. OpenAI describes this as reverting to the legacy saved memories system, and a Try improved memory link there takes you back.
- To leave both systems untouched for one conversation, use a Temporary Chat. Temporary Chats do not use existing memories and do not create new ones, and OpenAI states they are deleted from its systems after 30 days.
- Turning a setting off does not delete anything already stored. It stops new use and new writing, and that is all.
Before you decide anything is broken, check which layout your account has. A large share of confusing advice online describes the two-toggle screen to someone looking at the memory summary screen, or the reverse. Open Settings, Personalization, Memory first, see whether you are looking at switches or a summary, and only then follow instructions written for that layout.
Two settings was never the whole picture
Even at its peak, the two-toggle model was an incomplete map, because ChatGPT has other places where memory behaves by different rules. Projects are the main one. When you create a project you choose whether its memory is project-only or default, and project-only can only be set when starting a new project. Existing projects stay on default.
- Project-only memory: your previously saved memories are not referenced during those chats, and chats can reference other conversations within the same project. It is a walled garden with its own history.
- Project memory has no list. OpenAI states project memory does not show a list of memories the way personal memory does. To make it ignore something, delete that conversation or move it to another project.
- Sharing a project forces project-only memory from that point on, and it cannot be reverted. Shared projects have no access to any member's outside context, custom instructions or memories.
- Temporary Chat sits outside all of this. Project memory, personal memory and the summary are all bypassed.
So the real question is not two settings but four scopes: default memory, project-only memory, Temporary Chat, and memory switched off. Guides that stop at the two toggles cannot explain why a project chat suddenly forgets your dietary preferences. Project-only memory is the answer, working as designed.
What happens to each when you delete a chat
Deleting a chat takes it out of what reference chat history can draw on going forward, but it does not touch saved memories. OpenAI is direct about this: deleting a chat does not remove saved memory from that conversation, because those memories are stored separately from chat history. To fully remove a detail you have to delete the saved memory in Settings and the chat where you shared it.
On the updated system the cleanup list is longer, because memory is synthesised from more surfaces. OpenAI's guidance for fully deleting something ChatGPT may know about you is to delete every source where it appears, including past chats, archived chats, files, the memory summary, and to disconnect any connected apps that may hold that information. Archived chats are the one people forget: archiving is not deleting.
There is also a retention detail worth knowing. OpenAI states it may keep a log of deleted saved memories for up to 30 days for safety and debugging. And if you turn memory off, clear it, then turn it back on later, ChatGPT may build new memories from chats still sitting in your history, including old ones. A clean slate needs the chats gone, not just the memories.
The part most articles get wrong
Nearly every guide still presents the two toggles as the permanent shape of ChatGPT memory. That framing is now out of date. On 4 June 2026 OpenAI began rolling out a memory system where the two settings collapse into one memory control plus an editable summary, and it explicitly labels the saved memories system the legacy saved memories system. The rollout started with Plus and Pro users in the US and was set to expand to more countries and to Free and Go plans over following weeks.
Part of the confusion is not the writers' fault. OpenAI's own help centre currently disagrees with itself across pages. One article still walks you through the two toggles as though they are the live product, while the Memory FAQ files saved memories under a heading called Legacy memory system and leads with a single memory control and an editable memory summary. Both pages were updated in July 2026, so neither is stale. Treat the FAQ as the description of the current system and the two-toggle article as the description of what accounts that have not been migrated still see.
The two features also age differently. A saved memory from last year still reads as current fact, which is why OpenAI cited contradictions like training for a marathon sitting alongside a sprained ankle as a reason to rebuild the system. History-derived context avoids that trap but has its own: you cannot correct what you cannot see.
Which one should you leave on
The useful question is not which feature is better but which kind of forgetting you can tolerate. Pick based on what you actually use ChatGPT for.
- Both on: best continuity, least predictability. Good if you use ChatGPT for ongoing work and want it to keep up without being told twice.
- Saved memories only: ChatGPT knows the handful of facts you approved and nothing else. Good for shared devices, sensitive topics, or anyone who dislikes the surprise effect.
- Reference chat history only: rarely the right choice, because you lose the one layer you can actually inspect and edit.
- Neither: every conversation starts cold. Fine if you use ChatGPT for one-off questions, painful if you use it for anything continuous.
- Any of the above plus Temporary Chat for the occasional conversation you do not want feeding either system.
The gap neither setting closes
Both features only know what you typed into ChatGPT. The insurance PDF in your email, the whiteboard photo, the voice note from a call, none of it reaches either layer unless you retype it. That is why MemX exists: you save the photos, PDFs, voice notes and chats themselves, MemX reads them with OCR and transcription, and you ask in plain language and get the answer plus the source document. It is private by architecture, with per-user isolation, customer-managed encryption keys and encryption at rest. It is a place to keep things you need to look up, not a replacement for what ChatGPT does inside a conversation.
Whichever setup you choose, choose it deliberately. The difference between a list you can edit and a synthesis you cannot is the difference between knowing what an assistant knows about you and finding out by accident.
01Is reference chat history the same as saved memories?
No. Saved memories are specific details you asked ChatGPT to keep, stored as an editable list. Reference chat history is ChatGPT pulling relevant context out of past conversations on its own. They are separate settings and you can run either without the other.
02Why does ChatGPT know things I never asked it to remember?
Usually reference chat history, which reads past conversations without being asked. Sometimes it is a saved memory ChatGPT wrote automatically. Check your saved memories list first. If the detail is not there, it came from your chat history.
03Does deleting a chat delete the memory from it?
Only partly. Deleting a chat removes it from what reference chat history can use, but saved memories are stored separately and survive. OpenAI says you must delete both the saved memory in Settings and the chat where you shared it.
04Where is the reference chat history setting?
Profile icon, then Settings, then Personalization, under Memory. On accounts with the June 2026 update the two toggles are replaced by one memory control and a memory summary, and the old layout is reachable via Settings, Memory, Saved memories.
05Can I turn off reference chat history but keep saved memories?
Yes, on the two-toggle layout. OpenAI states you can turn on saved memories, both, or neither, and change it whenever you like. Turning either off does not delete what is already stored, it only stops further use.
