ChatGPT 記憶如快取、自訂指示如設定檔：規則應放在哪裏？

If you write software, you already know the difference between a config file and a cache. A config file is something you wrote, you can read it, and it behaves the same way every time. A cache is something the system fills in for itself, it changes as things change, and you should never trust it as your source of truth.

That is the cleanest way I have found to explain ChatGPT custom instructions versus ChatGPT memory to people who keep seeing ChatGPT "remember" things and are not sure where it comes from.

I checked the details below against OpenAI's Help Center in October 2026. Menus and limits vary by plan, so check your own Settings.

Custom instructions: the config file

Custom instructions are direct guidance you give about what ChatGPT should know and how it should answer. OpenAI says they apply to all chats and are available on all plans, on web, desktop, iOS and Android.

On the web or desktop app, open Settings, choose Personalization, make sure Enable customization is on, and type into the Custom Instructions field. On iOS and Android the menu is called Customize ChatGPT.

There is a size limit. Free and Go users can save up to 1,500 characters. Plus, Pro, Business, Enterprise and Education users can save up to 5,000.

Good things to put in there are stable preferences:

Your role
Your tone
"Ask me a question if a fact is missing."
"Never invent prices."

It is predictable. The same text goes in every time, and you can read exactly what is there.

Memory: the cache

When Memory is on, ChatGPT can remember relevant preferences and details from your chats and other sources. OpenAI lists those sources as past chats, saved memories, custom instructions, files in Library and content from connected apps such as Gmail, depending on your plan and region.

Two things matter if you use ChatGPT for work.

First, memory does not keep every detail of every conversation. ChatGPT decides what is relevant, and what it holds can change over time. That is cache behavior.

Second, you control it. Go to Settings, then Personalization, then Memory to review what it knows, correct it, or turn it off. Turning memory off does not delete your past chats.

OpenAI answers the question directly in its Memory FAQ: custom instructions are what you tell ChatGPT, and memory is what ChatGPT picks up.

Side by side

	Custom instructions	Memory
Who writes it	You	ChatGPT, from your chats and other sources
Predictable?	Yes, same text every time	No, changes with your context
Where to edit	Settings, Personalization, Custom Instructions	Settings, Personalization, Memory
Best for	Tone, role, standing rules	Helpful background that isn't critical
Size limit	1,500 or 5,000 characters	Not a fixed list you write

You can see both at work. When ChatGPT uses personal context, a Sources link can appear under the answer showing that a past chat, saved memory or custom instruction contributed. OpenAI notes Sources may not show every factor, so treat it as a hint.

My rule for business work

Anything you want every time goes in custom instructions. Memory is a convenience, not a database.

Do not count on memory for prices, dates, client names or policies. If a fact has to be right, paste it into the chat you are using. A cache can be stale, and a polite wrong answer is worse than a blank.

If you repeat the same background for one kind of work, a Project is better than either one. OpenAI says project instructions apply only inside that project and override your global custom instructions. So a bakery's price list and brand voice belong in a Project, not in your global settings where they follow you into every unrelated chat.

Temporary Chat and privacy

Before you start a Temporary Chat, you can choose whether it uses existing memories and custom instructions, or pick Unpersonalized to leave them out. OpenAI says temporary chats do not create or update memories. The choice is locked once the conversation starts, so decide first.

On personal accounts, OpenAI may use chats and remembered information to improve models when "Improve the model for everyone" is on. You can turn that off in Data controls. OpenAI also says custom instructions information may be used to improve model performance. By default it does not use content from ChatGPT Business, Enterprise or Edu workspaces to train its models.

Either way, keep client secrets, card numbers and health details out of custom instructions. Treat them like any other text field where you do not fully control the storage.

A five-minute cleanup

Here is what I would do today:

Open Settings, Personalization, Custom Instructions. Write three to five standing rules in plain words.
Open Memory and read what is in there. Delete anything wrong or anything you would not want in a screenshot.
Decide if memory should stay on. If more than one person shares a login, I would turn it off.
For any repeated kind of work, create a Project and put its rules there instead.
Before pasting something sensitive, use a Temporary Chat with Unpersonalized selected.

None of this takes long, and it ends the "why did it say that?" surprise.

Practice moving rules into a Project

I built a free lesson where you set up a ChatGPT Project with brand voice instructions, a price list and a style, in a practice copy of ChatGPT so nothing touches your real account: ChatGPT Projects: Keep Your Brand Voice, Files and Instructions in One Place.

Read the full guide: ChatGPT Memory vs Custom Instructions: What's the Difference