Memory and personalization in ChatGPT
Memory, chat history, custom instructions, personality, and control boundaries.
ChatGPT guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on chatgpt.com
Memory lets ChatGPT carry useful context across conversations. Personalization can also reflect your preferred tone, instructions, and past interactions.
This reduces repeated explanation. It also means an old, incorrect, or sensitive detail may quietly influence a future answer.
The kinds of context are different
| Context source | What it does | Main caution |
|---|---|---|
| Current chat | Uses messages in this conversation | Long chats can drift or contradict themselves |
| Saved memory | Retains specific details across chats | May be outdated, unwanted, or too broad |
| Reference chat history | Uses patterns or facts from past conversations | Less visible than a fixed saved memory |
| Custom instructions or personalization | Sets general preferences and background | Can distort tasks where the preference does not fit |
| Project context | Uses chats, files, and instructions inside a Project | Old project material can conflict with new work |
| Temporary Chat | Avoids normal history and memory behavior | Still uses cloud processing and temporary retention rules |
OpenAI separates saved memories from chat history. Review the current controls under Settings and OpenAI's memory documentation.
Good things to remember
Memory works best for durable, low-risk preferences:
- Preferred name or pronouns when you want them used.
- Writing style and reading level.
- Dietary preferences.
- Recurring project terminology.
- Device ecosystem.
- Goals that will remain relevant.
Avoid turning Memory into a medical chart, password vault, legal file, or complete life history.
A useful instruction is:
Remember that I prefer concise explanations with a practical example. Do not save health, financial, or family details unless I explicitly ask.
The system may not follow that boundary perfectly, so review it.
Audit memory regularly
Ask What do you remember about me? and inspect the result.
| Finding | Action |
|---|---|
| Incorrect detail | Correct or delete it |
| Temporary fact | Remove it after the project or event |
| Sensitive detail | Delete it and consider deleting the original chat |
| Vague inference | Replace it with a narrow explicit preference |
| Useful durable preference | Keep it and check whether it is applied appropriately |
Deleting a chat and deleting a saved memory are separate actions. To remove a detail fully, you may need to delete both the memory and the conversation where it appeared.
Personalization can create a bubble
An assistant that remembers your preferences may challenge you less. It can repeatedly frame advice around an old goal or reinforce one side of a personal story.
For decisions, ask:
Ignore my usual preference and make the strongest case for the alternative.Which parts of your answer came from memory rather than this prompt?What assumption about me most affects this recommendation?Answer as though this were our first conversation.
Use Temporary Chat or turn Memory off when you need a clean perspective.
Projects are not the same as personal memory
A Project can have its own files, chats, and instructions. This is often better than global Memory for work that belongs to one bounded context.
Keep personal details out of a work Project and work details out of a family Project. Name authoritative files clearly and remove obsolete versions.
For collaborative Projects, assume shared members may see project material. Review contents before inviting anyone.
Privacy and training controls
Memory and model-training controls are related but separate. If “Improve the model for everyone” is enabled on a consumer account, OpenAI says chats and saved memories may be used to improve models.
Turning training off does not delete memory. Turning Memory off does not necessarily delete existing saved memories or chats.
Use the current Data Controls FAQ rather than relying on old screenshots.
Design a memory policy for yourself
A simple policy can be:
- Save preferences, not secrets.
- Use Projects for bounded work.
- Use Temporary Chat for sensitive or one-off topics.
- Audit saved memory monthly.
- Correct errors immediately.
- Store important records elsewhere.
This creates useful continuity without treating ChatGPT as a permanent personal database.
When this is the wrong tool
Memory is the wrong tool for emergency information, medication schedules, legal records, passwords, confidential case histories, or facts that must be complete and exact.
Use a password manager, health record, calendar, database, or encrypted notes system designed for the job.
Memory should make conversation easier. It should not become the only place your life is organized.
