AI for Relationships
How to Clear ChatGPT Memory, and What to Save First
How to clear ChatGPT memory: delete one memory or all of it, make it actually gone, fix a full memory, and what to copy out before you wipe it.
To clear ChatGPT’s memory, open Settings, then Personalization, then Memory. Delete single entries from your saved memories, or clear the whole thing, and if you want a detail truly gone, also delete the chat where you first mentioned it. You can also type forget that I followed by the detail in any conversation.
That is the two-minute answer. The more useful question is what to delete, because the reflex is to bin the oldest entries, and that quietly throws away the wrong things. Memory settings have no export button, so anything you do not copy out first is gone from ChatGPT for good.
How to clear ChatGPT memory, step by step
OpenAI changes these labels often, and some controls only appear on some plans. The steps below match OpenAI’s own help page as of September 2026, and the path starts from Settings on the web and in the apps.
Delete one memory
- Open Settings, then Personalization, then Memory.
- Open Manage or Saved memories, whichever your account shows.
- Find the entry and delete it. On some plans the list can be searched and sorted, which helps once it runs past a screen or two.
The faster route is conversational: say forget that I moved to Denver and ChatGPT removes the matching entry. Check the list afterwards, since a conversational request is a sentence, not a receipt.
Clear all of it
If your account shows a memory summary (OpenAI’s newer, improved memory), open the ••• menu on it and choose Delete and turn off memory. That deletes what the summary holds and switches Memory off. Your past chats stay.
On the older saved-memories list, the option to clear everything sits in the same Manage view. It empties the list and leaves your chat history alone.
Make it actually gone
This is the step the one-button answers skip. ChatGPT’s memory now draws on several sources: saved memories, past chats, custom instructions, files in your Library, and connected apps such as Gmail. Deleting a saved memory stops it being used in future answers, but a detail that still sits in an old chat can be picked up again.
To remove something completely, delete it in each place it lives:
- the saved memory or the line in the memory summary
- the chat where you first shared it, including archived chats
- any file in Library that mentions it
- a connected app, if that is where it came from
Deletion takes time to propagate, and OpenAI says it may keep logs of deleted saved memories for up to 30 days for safety and debugging.
Stop it saving new ones
Turn off Reference saved memories under Settings, Personalization, Memory. That also turns off Reference chat history, and OpenAI schedules what it learned from past chats for deletion within 30 days. For a single conversation you want kept out entirely, start a temporary chat: it never creates or updates memories, and you can pick Unpersonalized so it ignores the memories you already have.
If ChatGPT says memory is full
ChatGPT keeps two different kinds of memory and only one of them fills up.
Saved memories are discrete entries: your job, your dietary preferences, the tone you want, the project you are shipping. This is the finite store. When it is full you get a notice telling you that memory is full and that you can forget existing memories to make space.
Chat history is the separate, much larger record of your conversations, which ChatGPT draws on through background synthesis. OpenAI says there is no separate storage limit for what it can reference this way, which is why the assistant can still feel like it knows you even while the saved-memories store is jammed.
The practical consequence is that “memory full” is not a signal that ChatGPT has learned too much about you. It is a signal that one specific drawer is out of room. On Plus and Pro, automatic memory management now moves less-used entries to the background to hold that moment off, but the drawer is still finite, and free accounts meet the ceiling first.
Before you wipe it, clear it in this order
Four steps. The order matters more than the steps do, because the destructive one is last and most people start there.
1. Read the list before you touch it. Open Settings, then Personalization, then Memory, and read the whole list or summary, not the first screen. If you would rather have it as text you can keep, ask in the chat: list everything you have saved about me, verbatim. There is no export button anywhere in the interface, so this is the only copy you are going to get.
2. Merge the duplicates. Ask ChatGPT to go through its saved memories, combine the ones that say the same thing, and drop anything that has expired. This reclaims more room than deleting entries one at a time, for a reason worth understanding, which is the next section.
3. Move the dated facts out. Anything about another person, anything with a date attached, anything you will need on one specific future occasion: write it somewhere it will still be findable in November. This is the step everyone skips, and it is the one that cannot be undone by changing your mind later.
4. Delete the rest. Remove entries individually, or say forget that I followed by the detail. Clearing everything is in the same place, and it empties the saved memories without touching your chat history, which is stored separately. Some plans now keep a version history of saved memories; without it there is no undo.
Where the space actually went
Most full stores are not full of a hundred different facts. They are full of one fact written six ways.
A new memory is not reconciled against what the store already holds. If your preference for short answers comes up in a Tuesday conversation, again while drafting an email in March, and again when you complain about a long reply, that can land as three entries in three phrasings. None of them is wrong. All of them are the same instruction, and each is occupying a slot.
Stale entries do the same thing more slowly. The half marathon you were training for in March is still in there in August. The client project you finished last quarter is still described in the present tense. Nothing removes an entry when the world moves on, because nothing knows that it has.
That is why consolidation beats deletion as a first move. Deleting one entry at a time frees one slot each. Merging a cluster of near-identical entries frees several at once and loses nothing, because the merged version still says what all of them said.
Why it fills up faster than it used to
Saved memories used to require a cue. You said “remember that I am vegetarian” and an entry appeared. Everything you did not explicitly flag was forgotten.
That changed. ChatGPT’s memory is now synthesized in the background, so context that comes up naturally in a conversation can be written without you asking. OpenAI’s stated goal for this is freshness: memory that updates itself as time passes, so “you are travelling to Singapore in July” becomes “you went to Singapore in July” once the trip is over.
For facts about you, that is a real improvement. The side effect is that the store now grows on its own, and it grows fastest during the periods when you are using the assistant most heavily. Most people discover the ceiling in the middle of a busy week, which is exactly when they have the least patience to curate it.
The controls, and what each one actually touches
Six controls govern this, and they are easy to confuse because several of them sound like they delete something and do not.
| Control | What it does | What it does not do |
|---|---|---|
| Reference saved memories, off | Stops ChatGPT using saved memories, and turns off chat history too | Does not delete the entries, and does not free space |
| Reference chat history, off | Stops it drawing on past conversations and schedules what it learned from them for deletion within 30 days | Does not delete the conversations |
| Don’t mention this again | Reduces future references to one detail | Does not delete the chat or file it came from |
| Temporary chat | A conversation that never creates or updates memory, and can ignore existing memory if you pick Unpersonalized | Does not change your settings for other chats |
| Project-only memory | Keeps a project’s context inside that project, and can now be switched on an existing project | Does not give you more room in the main store |
| Delete all saved memories, or Delete and turn off memory | Empties the saved memories or the memory summary | Does not delete your chats |
The useful pairing for most people is nothing turned off and one curation pass a quarter, with a temporary chat for the conversation you do not want landing in the store at all. Switching memory off entirely solves the full message by removing the feature, which is a fix in the same way that unplugging a fridge stops it beeping.
The failure worth knowing about
There is a reported behaviour that makes this more than a housekeeping chore: when the store is full, ChatGPT can still reply as though it has saved something new.
You say “remember that Priya moved to the Lisbon office”. It answers that it will remember. Nothing was written, because there was no room. You find out weeks later, when you ask about Priya and get nothing back.
The check takes ten seconds. After you ask it to remember something you actually care about, ask it to read the new entry back to you. If it cannot quote the entry, nothing was written and the confirmation you got was a sentence, not a receipt.
This is the part that should change how you think about the feature. A store that can silently decline a write is fine for preferences, because a dropped preference costs you one slightly-off answer that you immediately notice and correct. It is not fine for anything you are relying on months later, because by then the fact is gone and you have no record that it was ever refused.
What deserves the space
Once you accept the store is small and lossy, triage gets easy. Keep the things that improve every future answer, and move out the things you need to retrieve on a specific day.
| Keep it in ChatGPT’s memory | Move it somewhere durable |
|---|---|
| How you want responses formatted | What a client said their Q3 constraint was |
| Your role, your industry, your stack | Which colleague is quietly job hunting |
| Standing preferences and constraints | The introduction you promised to make |
| The project you are shipping this quarter | A friend’s diagnosis and when to ask about it |
| Your writing tone | Who you met at the conference and what they needed |
The left column is a description of you, and it stays useful no matter what you happen to be asking about. The right column is a set of dated facts about other people, each of which matters on exactly one future occasion. Storing the second kind in a small store that overwrites itself is the mismatch, and it is the reason people feel the memory is unreliable even when it is working as designed.
Run a real list through that and the verdicts come quickly. Open your saved memories and you might find:
- Solutions consultant at a mid-size logistics firm. Keep. It shapes every answer.
- Prefers short replies with no preamble. Keep.
- Wants concise answers. Merge into the entry above.
- Training for a half marathon in March. Expired. Delete.
- Client Marcus is worried about the Q3 renewal after a bad support experience. Move.
- Promised to introduce Nadia to a designer. Move.
- Colleague Tom is quietly interviewing elsewhere. Move, and think about whether it belongs in an AI account at all. That question is worth its own read: the people in your AI’s memory never agreed to be there.
Four verdicts, and only one of them needs a destination. The move pile is almost always the smallest by count and the largest by consequence, which is exactly why it should not be sharing a ceiling with your preferred bullet style.
Where the second column should live
Facts about people need three things the saved-memories store does not offer: room to grow without evicting anything, an index by person rather than by recency, and a guarantee that a write actually happened.
That is what a relationship memory app is. You write a note after a conversation, it files under the person, and it is still there in November. Nothing is competing for a slot, so nothing gets pushed out to make space for a formatting preference. For the wider comparison of what each assistant holds, see what ChatGPT, Claude, Gemini and DeepSeek remember. If hitting the cap is what has you looking elsewhere, Claude’s memory is readable entry by entry and takes a one-time import of your ChatGPT profile, and for the version of this question that starts from the workflow rather than the error message, see can ChatGPT be your personal CRM.
Keep using ChatGPT’s memory for what it is good at. It genuinely saves you re-explaining your role and your format preferences at the start of every session, and the writing you get back is better for it. Just stop asking it to be the only place a promise to a real person is recorded.
Key takeaway: Clear ChatGPT memory under Settings, Personalization, Memory, and delete the original chat too if you want a detail truly gone. “Memory full” means the finite saved-memories store is out of room. Clear it in order: read the list, merge duplicates, move the dated facts somewhere durable, then delete. It fills faster now because memory is written in the background, and a full store can silently decline a new write, so keep preferences there and keep dated facts about people somewhere built to hold them.
FAQ
How do I clear ChatGPT’s memory?
Open Settings, then Personalization, then Memory, and delete individual entries or clear them all. If your account shows a memory summary, use Delete and turn off memory from its menu. You can also tell ChatGPT in a conversation to forget a specific detail. Clearing memory does not delete your chat history, which is stored separately.
How do I delete ChatGPT memory permanently?
Delete the saved memory, then delete the chat where you first shared the detail, plus any Library file or connected app that holds it. OpenAI says deletion takes time to propagate and that it may keep logs of deleted saved memories for up to 30 days.
How do I see everything ChatGPT has saved about me?
The Memory page lists your saved memories, or shows a memory summary on the newer version. OpenAI notes the summary is a high-level view and can leave details out. There is no export button, so if you want a copy, ask in the chat for it to list everything it has saved about you verbatim, and paste the result somewhere you control before you start deleting.
What is ChatGPT’s memory limit?
OpenAI has not published a specific number of entries or characters. What is visible is the behaviour: the saved-memories store has a ceiling, and once you reach it you get a notice that memory is full and that you can forget existing memories to make space.
Can I increase ChatGPT’s memory limit?
There is no setting that raises the ceiling. The two things that genuinely buy you room are merging duplicate entries and taking the dated facts out of the store entirely, since those are the entries that accumulate fastest and pay off least often.
Does deleting a chat delete the memory from it?
No. Saved memories are stored separately from conversations, so a memory created during a chat survives that chat being deleted. To remove it you have to delete the memory entry itself, and to remove it completely you delete both.
Can I move my ChatGPT memory to Claude?
Partly. You can ask ChatGPT to list what it has saved about you and paste that into Claude’s import. What moves is a summary of your preferences; the context inside your old conversations stays behind. Does Claude have memory walks through what Claude keeps once it arrives.
Does a full memory stop ChatGPT from working?
No. Conversations carry on normally and it can still draw on your chat history. What stops is the writing of new entries, which is why the failure is easy to miss: nothing breaks, it just quietly stops recording.
What should I not store in ChatGPT’s memory?
Anything you need to retrieve reliably on a specific future date, and anything about another person that they would not expect to be sitting in an AI account. Both are covered in more depth in telling AI about other people.
Intriq keeps dated notes filed under each person, with nothing competing for a slot. Start with the personal CRM hub or the AI relationship assistant hub.