ChatGPT Memory Explained — What It Actually Remembers

Three Systems, Not One

Saved memories are specific facts ChatGPT keeps in an editable list — things you asked it to remember (“remember that I'm vegetarian”) or things it noticed and saved on its own. You can see the whole list and delete individual entries. Chat history (sometimes shown as “Reference chat history”) is different: instead of a fixed list, ChatGPT looks back across your past conversations and pulls forward whatever seems relevant to what you're asking right now. You can't inspect or edit it item by item — you can only turn the whole thing on or off.

As of June 2026, a third layer sits underneath both: a background process OpenAI calls Dreaming. It periodically rereads your conversation history and rewrites what it has learned about you, without you asking it to save anything — correcting stale facts the way a human would naturally update an impression of someone over time.

The List You Can Actually See and Edit

Saved memories are the one part of ChatGPT's memory you can open, read line by line, and prune.

Saved memories work a lot like Custom Instructions, except ChatGPT updates the list automatically rather than you managing it by hand. Say something worth remembering — explicitly (“remember that I work in marketing”) or in passing (“I'm allergic to shellfish”) — and ChatGPT may store it as a discrete entry. Go to Settings → Personalization → Memory → Manage to see the full list, delete anything wrong or outdated, or wipe it entirely. You can also just ask in conversation: “what do you remember about me?” or “forget that I mentioned Seattle.”

There is a practical ceiling here — the list has a size limit, and once it's full, ChatGPT can't add new entries until something is removed. A useful trick when that happens: ask ChatGPT to summarize and condense your existing memories into fewer, denser entries, which frees space without losing the substance.

The Layer You Can Only Toggle, Not Edit

Reference chat history is a separate setting that lets ChatGPT draw on the substance of your past conversations — not a stored list, but a live look-back across what you've discussed before. If you once mentioned you're training for a marathon, it might factor that in weeks later when you ask for a dinner suggestion, even though you never explicitly asked it to remember that fact.

The key practical difference from saved memories: chat history isn't a fixed record. OpenAI describes it as something that can shift over time as the model updates what's actually useful to keep in view, which means you can't open a list and see exactly what it's drawing on the way you can with saved memories. If something from your history keeps surfacing in ways you don't want, the available lever is coarse — turn chat history off entirely from Settings → Personalization, rather than removing one specific detail.

The Newest Layer, Running Quietly in the Background

In June 2026, OpenAI began rolling out a system it calls Dreaming — described in its own announcement as a more scalable way to keep your profile accurate across years of conversations, instead of letting old facts pile up and go stale. The name is a deliberate echo: it's a background process that reviews your history and updates what ChatGPT knows about you the way sleep is thought to consolidate memory in humans.

OpenAI's own example: a memory like “you're going to Singapore in July” automatically becomes “you went to Singapore in July 2026” once the trip has passed, without you correcting anything yourself. OpenAI reports this pushed its internal factual-recall evaluation from 41.5% to 82.8% — numbers worth noting come from OpenAI's own unreleased benchmark and haven't been independently verified, so treat them as a company-reported figure rather than an audited one. Dreaming rolled out first to Plus and Pro users in the US.

What You Can See, Edit, and Turn Off

Turning memory off doesn't retroactively delete what's already been saved — it only stops new memories from forming going forward. To actually erase what's stored, use “Manage memories” and delete or clear the list directly. And deleting a conversation doesn't remove any memory that conversation already produced; the memory and the chat it came from are stored separately.

If Your Settings Screen Looks Different From a Friend's

Since June 4, 2026, OpenAI's own documentation refers to the old two-toggle setup (a separate switch for saved memories and one for chat history) as the “legacy memory system.” The current default on most accounts is a single, unified Memory control with one editable memory summary rather than two separate lists. Both layouts are still live depending on your account and rollout stage, which is why two people comparing settings screens can see genuinely different menus — it isn't a bug, it's a staggered migration.

If you're on the newer layout, Settings → Personalization → Memory shows a single summary that updates automatically as you chat, with a timestamp showing when it last changed (“updated 2 hours ago”). If you still see the older paired toggles, you can switch back to the legacy saved-memories view from within that same settings screen if you prefer the more granular list.

Keeping Memory Useful Instead of Noisy

Memory is genuinely useful for repetitive, ongoing work — not having to re-explain your job, your writing style, or your dietary restrictions every session saves real time. But it can also drift: a fact from a single old project can quietly color unrelated conversations months later. A quick health check worth doing occasionally: ask ChatGPT directly “what have you learned about my preferences from past chats?” and look for anything that's outdated or was never actually true, then correct it in conversation or delete it in Settings.

For any conversation you don't want factoring into your profile at all — testing a prompt, asking something unrelated to your usual context, anything sensitive — use Temporary Chat. It reads no existing memory and writes none, which makes it the cleanest way to get a response unaffected by everything ChatGPT has picked up about you so far.

Frequently Asked Questions