Memory is not a brain. It's notes plus retrieval.
The new system — OpenAI calls it Dreaming — runs in the background, reads across your conversations, writes notes about you, and quietly injects the "relevant" ones into every new chat.
So when ChatGPT acts weird, it's one of three failures: a stale note, a wrong note, or the right note pulled at the wrong moment. Every step below fixes one of those. That's the whole guide.
Find out what it thinks it knows
Open a new chat and paste this:
The memories are the boring part. The assumptions are where it gets uncomfortable — you'll see the inferences it's been quietly building, like deciding what you do for a living based on one question you asked eight months ago.
Then cross-check the official version: Settings → Personalization → Memory. The new memory summary is readable and editable — new with the June update, and where most of the fixes below happen.
Power-user move: memories get injected into chats as a hidden block, and you can make ChatGPT print the raw notes — timestamps and all:
Seeing the dated entries is the whole point. "[2024-11-03] User is preparing for a move" — that note might still be steering your answers today.
When an answer feels off, find the memory that caused it
This is the move almost nobody knows about, because the feature is days old.
ChatGPT now shows memory sources — the specific stored info it used to personalize a given answer. So when a response feels weirdly specific, oddly preachy, or just off: don't argue with it. Open the sources, find the memory that steered it, and delete that memory on the spot.
I debug AI retrieval at work all day, and this is the exact same workflow: don't fix the output, fix the input. One bad note can poison a hundred answers. Now you can find it in ten seconds.
Delete these four kinds of memories
- Dead projects. The wedding speech, the one-time trip, the apartment you didn't take. ChatGPT doesn't know the project ended — it keeps optimizing for a life event that's over.
- Emotions saved as facts. You vented about your manager once in March. The note says you have a bad relationship with your manager. It's still coloring your career answers in June. Your March self should not be writing your June answers.
- Other people's stuff. You helped a friend with their resume, planned your mom's birthday, drafted someone else's breakup text. ChatGPT can file all of it under you.
- Expired you. Old city, old job, old goals. The most dangerous category, because the notes were true once — they just aren't anymore.
The deep clean (10 extra minutes): copy your entire memory list into a text file, start a fresh chat, paste it in, and say:
Review its proposal, wipe your memory, re-add only the keepers. Yes — you're using the AI to debug its own notes about you. It's good at it.
Stop letting it guess. Tell it.
Everything so far was defense. This is offense. Instead of letting Dreaming infer who you are from scattered conversations, plant the memory you want it to have:
Five minutes, and you've replaced months of guesswork with a profile you wrote yourself.
Tell it when things are allowed to come up
The genuinely new capability, and my favorite part of the update: the memory summary now accepts instructions about when to bring things up. Not just facts — rules.
Work and personal in the same account, finally separated. Two sentences in the summary page and your AI stops mentioning your side project while you're drafting something for your boss.
Some things should never become a note
Gift planning. Salary negotiation drafts. Venting. Health questions you don't want resurfacing in three months. Use a temporary chat for those — nothing from it enters memory. If something already slipped in mid-conversation, just say "forget that" and confirm it's gone from the list.
One more thing, since memory is now on for free accounts: a lot of people have been having permanent-feeling conversations without knowing anything was being kept. Now you know.
Back it up. Seriously.
Users have reported their entire saved-memory panel going blank — every device, no warning, no undo. And you can't paste memories back in. If they vanish, you're re-adding each one by hand.
So: open your memory list, select everything, paste it into a note file, date it. Thirty seconds. It's the only insurance that exists.
The 5-minute monthly tune-up
Memory hygiene isn't one-and-done, because you keep changing and the notes don't know that. Once a month — or any time something big shifts (new job, new city, new goal):
- Re-run the Step 1 audit prompt
- Delete what's expired
- Update your Step 4 seed
- Re-copy the backup