Open the controls
Go to Settings → Personalization → Memory. You'll see separate toggles for "Reference saved memories" and "Reference chat history," plus a way to manage the saved-memories list.
The two memory layers
Saved memories vs reference chat history.
ChatGPT's memory is really two systems. One is a tidy, editable list of explicit facts you can read and prune. The other, launched in April 2025, is an invisible recall of your past chats — no list, just an on/off switch. Knowing which is which tells you exactly what you can audit, what you can't, and how to turn each off.
Quick answer
One you can see, one you can't
Saved memories are explicit, observable facts. Reference chat history is an implicit layer that has no editable list — you can only toggle it. They're both controlled in the same place, but they behave very differently.
| Saved memories | Reference chat history | |
|---|---|---|
| What it stores | Explicit facts — name, role, tone, preferences. | Implicit patterns from your past conversations. |
| Can you see it? | Yes — an editable list. | No — no visible list, only a toggle. |
| How it's updated | Automatically, or when you say "remember this." | Continuously, in the background. |
| Launched | Available earlier as the editable memory list. | April 10, 2025. |
| Control | "Reference saved memories" toggle + Manage memories. | "Reference chat history" toggle. |
Both live in Settings → Personalization → Memory. Per OpenAI, turning off "Reference saved memories" also turns off "Reference chat history." Confirm on OpenAI's Memory FAQ.
How to manage each
Auditing and controlling the two layers
The two layers need different handling — one you prune item by item, the other you mostly switch off.
Audit saved memories
Open "Manage memories" to read every stored fact. Edit anything that's wrong, delete anything you don't want kept, or clear them all. This is the only layer you can inspect item by item.
Handle chat history
Reference chat history has no list to audit. To reduce what it can draw on, turn the toggle off and consider deleting old conversations you don't want reflected in future chats.
Go memory-free when needed
For a one-off private session that ignores both layers, use Temporary Chat instead of changing your settings — it won't read or write either layer.
Related: how memory works · memory vs custom instructions · memory limits · delete memory
Key behaviors
Things people get wrong about the two layers
| Assumption | Reality |
|---|---|
| "Turning off saved memories deletes them." | It doesn't. The list stays until you delete it. See delete memory. |
| "Chat history is just another editable list." | No — it has no visible list, only a toggle. You can't read what it draws on. |
| "I can leave chat history on after disabling saved memories." | Per OpenAI, turning off "Reference saved memories" also turns off "Reference chat history." |
| "Saved memories are the same as custom instructions." | Similar effect, different management. See memory vs custom instructions. |
If the invisible layer worries you
You can't audit chat history — but you can spread your context
Reference chat history is convenient precisely because it works silently, but that's also why it's hard to control. If you'd rather not concentrate all your context inside one assistant's invisible recall, a multi-model workspace lets you choose which model sees which prompt while you keep sensitive topics in memory-free sessions.
All-in-one
MultipleChat AI
multiplechat.aiUse ChatGPT, Claude, Gemini and other models in one place — choose where each prompt goes instead of feeding everything to a single assistant.
Privacy-first: MultipleChat doesn't save your chats to memory and doesn't share your data with model providers or let them train on it. Other privacy-focused options include Duck.ai, Proton Lumo, Brave Leo and Kagi Assistant; for full local control, self-hosted tools like Ollama or Jan keep everything on your device.
OpenAIChatGPT
chatgpt.comOffers Temporary Chat (no history, not used for training), a training opt-out under Settings → Data Controls, and editable memory. By default, consumer chats may be used to improve models unless you opt out — verify on OpenAI's help center.
AnthropicClaude
claude.aiAnthropic gives you data and training controls and has historically taken a privacy-conscious stance on consumer chats, but its training and retention settings have changed over time — check your current privacy settings in the app.
GoogleGemini
gemini.google.comGemini Apps Activity lets you pause history and delete past activity, but sampled conversations can be reviewed by humans, so avoid sharing confidential information. Review your Activity controls.
MicrosoftCopilot
copilot.microsoft.comMicrosoft 365 Copilot (business) doesn't use your organization's data to train models and adds Enterprise Data Protection; consumer Copilot has a privacy dashboard to review and delete your history. Check current Microsoft policies.
PerplexityPerplexity
perplexity.aiPerplexity has an "AI data retention" setting you can switch off so your searches aren't used to improve its models, plus account-level data controls. Confirm the current options in settings.
All guides
ChatGPT memory: every guide
How memory worksThe technical walkthrough.
Saved vs chat historyThe two layers compared.
Turn memory on/offManage the settings.
Delete memoryClear individual or all.
Memory & privacyRisks and how to reduce them.
Memory & trainingIs it used to train models?
Temporary ChatMemory-free conversations.
Memory not working?Troubleshooting steps.
Memory limitsHow much it can hold.
Team & EnterpriseMemory and data controls.
Memory vs custom instructionsWhat's the difference?
Memory FAQEvery common question.
FAQ
Saved memories vs reference chat history — answers
What each layer is, what you can see, and how to control them separately.
What is the difference between saved memories and reference chat history?
Saved memories are explicit facts ChatGPT stores about you — like your name, role and preferences — and they appear in an auditable list you can view, edit and delete. Reference chat history is an implicit recall of patterns from your past conversations; it's not a visible, editable list, only a toggle. In short: saved memories are observable and manageable item by item, while reference chat history works in the background. Both are controlled in Settings → Personalization → Memory.
What are saved memories in ChatGPT?
Saved memories are explicit, durable facts ChatGPT keeps about you — for example your name, your job, your tone preferences or that you're vegetarian. They're updated automatically as you chat, or when you tell ChatGPT to "remember this." They behave like auto-managed custom instructions, and you can open the list to view, edit or delete each entry.
What is reference chat history in ChatGPT?
Reference chat history, launched on April 10 2025, lets ChatGPT draw on patterns and context from your past conversations to make new chats more relevant. Unlike saved memories, it's not exposed as a list you can read or edit — you can only turn it on or off. It's an implicit layer rather than an auditable one.
Can I see and edit both layers?
Only partly. You can open the saved-memories list to see, edit and delete each explicit fact. Reference chat history has no visible list — there's nothing to read or edit item by item, just a toggle. So if you want to know exactly what ChatGPT "knows" about you, only the saved-memories layer is observable.
Where do I control saved memories and chat history?
Both are in Settings → Personalization → Memory. There you'll find separate toggles: "Reference saved memories" and "Reference chat history." You can turn them on or off independently, and open "Manage memories" to edit the saved-memories list. Verify the current layout on OpenAI's help center, as the interface changes.
If I turn off saved memories, does chat history stay on?
According to OpenAI, turning off "Reference saved memories" also turns off "Reference chat history." So disabling saved memories doesn't leave the chat-history layer running on its own. If you want a memory-free session without changing your settings, use Temporary Chat instead.
Does turning off a layer delete what is stored?
No. Turning off saved memories stops ChatGPT from using your stored facts, but it doesn't automatically delete them — they remain until you delete them in "Manage memories." Turning off reference chat history stops that implicit recall going forward. To actually erase explicit memories, delete them; to reduce the chat-history layer, you can also delete old conversations.
Which layer should I be more careful with for privacy?
Both matter, but reference chat history is harder to audit because it has no visible list — you can't see exactly what it draws on. Saved memories are more transparent because you can read and prune them. For sensitive topics, the cleanest option is Temporary Chat, which ignores both layers. Review your saved memories regularly and verify behavior on OpenAI's help center.
How does each layer reach the model?
Both layers work at inference time. Relevant saved memories and signals from past chats are injected into the model's context window alongside your prompt, like a hidden system note, so the model conditions its reply on them. The model isn't retrained per user. The difference between the layers is in what's stored and whether you can see it, not in how it's delivered to the model.
Why does ChatGPT remember something I never explicitly saved?
That's usually the reference-chat-history layer at work. Because it draws on patterns from past conversations rather than a fixed list, ChatGPT can reflect things you discussed earlier even if they were never written to saved memories. If this is unexpected, turn off reference chat history and consider deleting the relevant old conversations.
Are saved memories the same as custom instructions?
They're similar in effect but different in management. Saved memories are auto-managed and grow over time, behaving like custom instructions that ChatGPT writes for you. Custom instructions are a separate feature you write and edit yourself, and they're always applied. You can use memory and custom instructions together. See our guide on memory vs custom instructions for the full comparison.