Memory is the feature that makes ChatGPT feel like it knows you — until it remembers something wrong from three months ago and silently uses it on a project where it does not apply. Treat Memory like a small, leaky notebook you do not fully control: useful for stable preferences, terrible for anything that changes or anything important. This guide shows what to put in Memory, what to put in Custom Instructions instead, and how to keep both clean.
What this covers
How ChatGPT Memory actually works, the split between Memory and Custom Instructions, and the maintenance habit that keeps your stored memories useful instead of polluted. Also covers when to use Temporary Chats to bypass Memory entirely.
Key tools and concepts:
- ChatGPT: OpenAI’s conversational AI assistant — the product that brought the GPT models to a mass audience.
- Memory: A growing list of facts the model tries to retain across chats. Saved to a manageable list, but the model decides what to save unless you tell it.
- Custom Instructions: Two static fields you control. Survive forever, never get rewritten by the model.
- Temporary Chat: A mode that does not write to Memory or chat history. Use for one-off sensitive queries.
Who this is for
Anyone who has noticed ChatGPT “remembering” something across chats — for better or worse — and wants to control what it retains. Especially useful if you share an account context across multiple projects or roles.
When to reach for it
You are using ChatGPT regularly for two or more different contexts (work + personal, two clients, two projects) and the model is mixing them up. Or you have noticed it citing a fact about you that is no longer true.
Step by step
- Open Settings, then Personalization, then Memory. Confirm it is on. If you want zero retention, turn it off and skip the rest.
- Read your existing memories. Most users have 50-200. Skim for outdated, contradictory, or wrong entries.
- Delete anything you would not want a stranger to see — Memory can leak into responses unpredictably.
- Save 3-5 stable facts deliberately. Phrase: “Please remember that I am a Cornell grad student in CS.” The model echoes back “I will remember” — that is your confirmation.
- Move stable preferences to Custom Instructions instead. “Always answer in concise bullets” belongs there, not in Memory.
- For sensitive or experimental chats, open a Temporary Chat. Nothing is saved to Memory or your history.
- Review Memory monthly. It grows silently — without pruning, contradictions accumulate.
Memory vs Custom Instructions — when to use which
| Fact about you | Lives in |
|---|---|
| Your job, location, expertise level | Custom Instructions |
| Preferred answer style (length, tone) | Custom Instructions |
| Hard rules (“never write in marketing-speak”) | Custom Instructions |
| Current project, this week’s focus | Memory |
| Evolving interests | Memory |
| One-off task context | Neither — put it in the prompt |
| Sensitive personal info | Neither — do not save it |
First-run exercise
- Open Memory settings. Screenshot the list before you start — so you can compare in a month.
- Delete 30% of entries. Outdated, redundant, or just noise.
- Write 3 sentences for Custom Instructions: your role, your preferred answer style, one hard rule.
- Start one fresh chat. Notice whether the model picks up your Custom Instructions correctly. If not, adjust the wording.
Quality check
- Does your Memory list contain anything that contradicts something else in the list? Pick one and delete the other.
- Are there facts in Memory that should be permanent (Custom Instructions) or one-off (just the prompt)? Move them.
- Did you ever say something in frustration (“ugh, you always do X”) that the model misinterpreted as a preference? Find and delete it.
How to reuse this workflow
- Calendar a 10-minute Memory review on the first of each month. Otherwise you will not notice rot.
- Keep a personal note of your “canonical” Custom Instructions text. The UI is small and editing in place is annoying — drafting offline is faster.
- For multi-context users, accept that Memory is one shared bucket. If contexts are wildly different (medical research + game dev), use separate Workspaces or accounts.
Recommended workflow
Set 3-5 long-term facts deliberately. Use Custom Instructions for stable, structural facts (job, style, hard rules). Use Memory for evolving facts (current project, recent interests). Use Temporary Chats for anything you do not want retained.
Common mistakes
- Expecting Memory to retain everything you said in a chat — it does not, it selectively saves what the model deems important.
- Using Temporary Chats for things you actually wanted remembered. The chat history will not be there next week.
- Treating Memory like a personal notes app. It is not searchable, not editable in place, and the model can paraphrase or drop entries.
- Letting Memory grow past a few hundred entries without pruning. Stale entries crowd out fresh ones.
- Saving project-critical info in Memory instead of Project files. A Project file is yours; Memory is approximate.
- Sharing accounts and assuming Memory is private to “your” chats. It is account-level.
FAQ
- Can I export my Memory?: Not directly; you can read the list in Settings and copy manually. There is no official export.
- Is Memory used to train models?: Depends on your data settings. Check Settings, Data controls. Enterprise and Team plans default to off.
- Why did it forget something I told it last week?: The model decides what to save. If you said it casually, it may have skipped saving. Say “Please remember that…” explicitly for critical facts.
- Memory in Custom GPTs?: Custom GPTs generally do not access your personal Memory. They have their own system message.
- How do I delete everything?: Settings, Personalization, Memory, Manage. There is a “Clear all” option. Custom Instructions are separate.