IntermediateLesson 11 of 305 min read

Custom instructions and memory

Teach ChatGPT how you like to work with custom instructions and memory, and keep both tidy, useful and private.

In this lesson you will

  • Write concise custom instructions
  • Understand how memory differs
  • Review and delete memories
  • Avoid storing sensitive information

About 15 minutes

From a helpful stranger to a colleague who knows you

Every time Mina opens a new chat she types the same three lines: “I run a small bakery. Keep answers short. Use British spelling.” It is a small chore, but it adds up across a hundred chats, and some days she forgets and gets a long, American-spelled answer for a different kind of business. The fix is to teach the assistant once. Custom instructions and memory let ChatGPT carry your preferences and background into every conversation, so it behaves less like a helpful stranger and more like a colleague who knows how you work. This lesson shows how to set them up well, and, just as important, how to keep them under control.

Good to know

Names, menus and availability of these features vary by plan, region and app version, and change over time. The ideas below apply either way. If you cannot find a setting, check the help pages on OpenAI’s website.

Two different things

Custom instructions Memory
What it is Text you write that applies to your chats Facts ChatGPT may save from your chats
Who writes it You, on purpose You, or the assistant when you mention something worth keeping
Good for Style, format, role, rules Background about you, ongoing projects, preferences
Control Edit or clear any time View, delete single items, clear all, or switch off
Risk Out-of-date rules It remembers things you forgot you said

Writing strong custom instructions

Most people write instructions that are too vague (“be helpful”) or too long. A good set is short, specific and about how to answer rather than who you are in detail. Think about three questions:

  1. What should it know about me that changes the answer? Your job, your level, your location (country) for spelling, units and laws.
  2. How should it respond? Length, tone, format, level of detail, whether to ask questions first.
  3. What rules must it never break? No invented facts, say when unsure, no flattery, always show your working.
What it should know about me:
I own a small online shop selling handmade candles in Pakistan. I am not technical.
I write for customers who are mostly women aged 20 to 40.

How to respond:
Use short sentences and plain words. Give the answer first, then a brief reason.
Use British spelling and rupees. Use bullet points for lists.
If my request is unclear, ask one question before answering.

Rules:
Do not invent facts, prices or sources. If you are not sure, say so.
Do not flatter me. Point out weaknesses in my ideas politely.
The anti-flattery line

Chat assistants can lean towards agreeing with you. A line such as “point out weaknesses politely and do not flatter me” gives you more useful feedback on your plans and writing.

What not to put in instructions

  • Passwords, bank details, ID numbers or anything secret. Instructions are stored.
  • Private details about other people.
  • Long life stories. Keep it to what changes the answers.
  • Contradictions such as “be very brief” and “explain everything in detail”. It will guess which one to follow.
  • Temporary details such as “I am travelling this week”. Put those in the chat instead.

Memory: useful and a little spooky

With memory switched on, ChatGPT may save facts you mention, such as “I am vegetarian” or “my daughter is starting school”, and use them in later chats. This can be very helpful: recipe ideas that fit your diet, a reminder of your project, advice pitched at your level. It can also surprise you if you forget what you told it.

What do you remember about me? List everything, and tell me how I can delete or change it.

Asking this occasionally is a good habit. You can also tell it directly:

Please remember that I prefer metric units and that my shop name is Glow Candles.
Please forget that I told you about my daughter's school.
Good to know

To be sure something is really deleted, also check the memory settings page. Asking in chat is convenient, but the settings screen is where you can see the full list.

Using memory wisely

  • Review it monthly. Delete anything out of date or too personal.
  • Keep sensitive topics out: health, finances, legal trouble, other people’s secrets.
  • Use temporary chat when you want a clean slate with no memory used or saved.
  • Separate contexts. If you use one account for work and personal life, tidy memory or use projects (next lesson) to keep subjects apart.
  • Remember it is still stored. Convenience is not the same as secrecy.

When memory hurts

Old facts can leak into the wrong answer. If you once said you were writing a thesis on climate and now ask for party ideas, a good assistant should not mention it, but sometimes it does. If an answer seems oddly influenced by your past, ask: “Are you using anything from my memory in this answer?” and turn memory off for that chat, or start a temporary chat.

A tune-up exercise

  1. Open your personalisation settings and read what is there.
  2. Cut it down to the three questions above.
  3. Test it: ask the same question in a new chat before and after changing the instructions.
  4. Check memory and delete two things you do not need.
  5. Write a one-line reminder in your notes: “Review ChatGPT settings first Friday of the month”.
Common mistake

Writing instructions once and forgetting them. Six months later you change jobs or move country and the answers still assume your old life. Make a reminder to review.

Common mistake

Treating memory as a private diary. It is stored on an online service under your account, and the same privacy lessons apply.

Check what you learned

Five quick questions. Pick an answer and you will see straight away whether it is right and why.

What is the main benefit of custom instructions?

  1. The screen gets bigger
  2. You do not repeat your preferences every chat
  3. Chats are free
  4. It writes code

They carry your style and background into every chat.

Which belongs in custom instructions?

  1. Your bank PIN
  2. A friend’s private secret
  3. Your password
  4. How you want answers formatted

Keep to style and rules, never secrets.

How can you see what memory has saved?

  1. Ask in chat and check the memory settings
  2. It is impossible
  3. Restart the phone
  4. Delete the app

You can review, edit or delete memories.

Which option gives a clean slate with no memory used?

  1. A longer prompt
  2. Capital letters
  3. Temporary chat
  4. Regenerate

Temporary chats are designed not to use or save memory.

Why add a line like ‘do not flatter me’?

  1. To shorten answers
  2. To use images
  3. To save money
  4. To get more honest feedback

It nudges the assistant to point out weaknesses.

Summary

  • Instructions describe how to answer; memory stores facts.
  • Short, specific and rule-based instructions work best.
  • Review settings regularly.
  • Keep secrets and sensitive details out.

Next lesson

Next: projects, keeping each area of your work organised with its own context.

Questions

Do custom instructions work in every chat?

Generally yes, unless you use a mode or chat type that ignores them, such as a temporary chat on some versions. Check your settings page.

Can memory be wrong?

Yes. It can save a misunderstanding or an out-of-date fact. Review it and delete anything incorrect.

Should I switch memory off?

It is a personal choice. If you value privacy more than convenience, switch it off and use projects or pasted context instead.

Progress is saved in this browser.