Memory & instructions: teach ChatGPT to know you.
ChatGPT has two ways of remembering you: Custom Instructions, which you write, and Memory, which it learns on its own. Used well, they make it feel like it knows you. Used badly, they quietly make every answer worse. Here's what belongs where — and how to clean up Memory when it drifts.
One is a contract. One is an observation log.
They both shape future responses, but they work differently and need different upkeep.
What you explicitly tell ChatGPT about you and how to respond. Lives in Settings → Personalization, in two boxes.
What ChatGPT picks up from your chats and stores on its own. Helpful, but it can save things that are wrong, stale, or that you never meant it to keep.
Clean up what ChatGPT saved about you.
Here's your Memory panel. For context: you write in TypeScript these days, you run a weekly newsletter, and you often do want depth — not just quick answers. Review each item and decide: Keep the accurate, useful ones; Delete what's wrong, outdated, or sensitive.
Instructions, Memory, or neither?
Three pieces of context. Put each in the right home.
Write Custom Instructions that work.
Settings → Personalization gives you two boxes — "what to know about you" and "how to respond." Fill both with process and style, never secrets.
Memory is on by default. Switch it off for:
- Shared accounts (family/team) — memories get tangled.
- Sensitive sessions — PHI, confidential financials, NDA-covered work.
- Prompt tests where you want a clean baseline.
Turning Memory off doesn't disable Custom Instructions — the contract still applies.
Temporary chat stopped being all-or-nothing.
Temporary chat used to be one switch: a clean room with no memory, no Custom Instructions, no plugins, and nothing kept. OpenAI has now split that into a choice you make when you open the chat — and a second choice you make at the end.
No memory, no Custom Instructions, no plugins. Creates no new memories. This is the old behaviour, unchanged.
Can read your memories, Custom Instructions and plugins to tailor the answer. Still creates no new memories, and still stays out of your history.
Two details worth committing to muscle memory. First, the mode is set only when you start the conversation — you cannot flip it once you are typing. Second, you can now save a temporary chat into your history, which converts it into an ordinary chat: from that moment it follows your normal account-level personalization and model-improvement settings, and it can be referenced later like any other conversation.
“Temporary chat” is no longer a single guarantee, so the sensitive-session advice needs one extra beat: glance at which mode you opened before you paste anything you would not want personalized. Non-personalized is still the default, so the safe path is unchanged — but personalized temporary chats do pull your existing memory in, and a chat you save stops being temporary in every sense.
Useful side effect: personalized temporary chat is now the cleanest way to test a prompt against your real Custom Instructions without polluting Memory with the result.
Audit and rewrite tonight
Two tasks, 15 minutes. (1) Open Settings → Personalization → Memory and read everything stored about you — delete what's wrong, redundant, or shouldn't be there. (2) Rewrite your Custom Instructions with the two-box template above. Save, then test the difference on your next three prompts.
What you can do now
- Tell Custom Instructions (the contract) from Memory (the observation log)
- Write Custom Instructions with the two-box template
- Add to Memory deliberately with "Remember that…" — but keep critical context in Instructions
- Audit Memory monthly: delete what's wrong, outdated, or sensitive; keep accurate-useful
- Know when to turn Memory off entirely