How to Delete ChatGPT Memory, and Check That It Worked
By Matthew Firth, founder of ModelBrain · Oct 26, 2026 · 4 min readA delete is only as good as the check you can run afterward, so here is our own proposal: the two-query check, which asks the original question and then asks for the exact string, and counts the delete as done only when both miss. On October 26, 2026 we saved a memory in ModelBrain at 15:27:25 UTC, forgot it seven seconds later, and the same query returned 0 hits while the exact-string query returned three unrelated memories (top relevance 0.571).
Delete the saved memory
Open ChatGPT, then Settings > Personalization > Memory > Manage. Each saved memory has its own delete control, and you can also type "forget that I live in New York" into a chat. In the memory summary, the more-options menu (the three dots) offers "Delete and turn off memory", which clears what the summary shows and switches Memory off. We read these steps on OpenAI's Memory in ChatGPT page on October 26, 2026.
Delete the chat it came from, too
OpenAI says saved memories are stored separately from chat history, so deleting a chat does not necessarily delete a memory made from it. To remove a fact, delete the memory and the chat where you first shared it. The same page lists the other places a fact can live: archived chats, files in your Library, and connected apps.
Turning things off is not deleting
Switching off "Reference chat history" schedules what ChatGPT remembered from past chats for deletion within 30 days, and your original chats stay until you delete them. It does not delete saved memories. OpenAI also says it may keep logs of deleted saved memories for up to 30 days for safety and debugging, and that deletions can take time to propagate.
The two-query check
The settings screen tells you a memory is gone. It cannot show you that recall agrees. Open a new chat and ask two things. First, ask the question the memory used to answer ("what is my staging database hint?"). Second, ask for the exact words ("do you have anything containing Biscuit-Quartz?"). Count the delete as done only when both come back empty. A miss on the first query alone can mean the answer was reworded or ranked lower, so the exact-string query is the one that catches a leftover.
The same check, timed, on ModelBrain
We ran it against a local ModelBrain 0.6.1 store on October 26, 2026. The memory was a made-up line containing the string Biscuit-Quartz-7741.
| Step | Time (UTC) | Result |
|---|---|---|
| Save the memory | 15:27:25 | Stored under a key |
| Recall by topic | 15:27:2x | 1 hit, relevance 0.833 |
| Forget it (default mode) | 15:27:32 | Seven seconds after saving, key tombstoned |
| Recall by the same topic | right after | 0 hits |
| Recall by the exact string | right after | 3 hits, all unrelated, top relevance 0.571, none containing the string |
| Open the key directly | right after | Text still readable, flagged deleted: true with the deletion time |
That last row is the part worth knowing. A tombstoned memory disappears from recall at once, but the text is cleared for good only after 30 days (our FAQ gives the same figure), and for the first 15 minutes you can undo the forget from the menu. OpenAI describes a 30-day window for logs of deleted memories too. The difference we can offer is that you can see ours: opening the key shows the deleted flag and the timestamp, so "gone from recall" and "not yet purged" are two states you can tell apart.
Why we call it a check and not a guarantee
Our forget-guarantee article describes the modes behind that call. The check above is the user-side half: whatever tool you use, the delete is finished when the original question and the exact string both miss. If you want one memory your assistants share and you can inspect, see Claude memory vs ChatGPT memory.