Why ChatGPT forgets, and what to do about it

“ChatGPT forgot what I told it” usually means one of four things, and they have four different fixes:

Here is why the thread forgets, and which of the four fixes changes the answer.

1. The thread got long, and the earliest turns fell out of the working context.
2. You started a new chat, which by design does not carry the old one.
3. You expected saved memory to hold a detail, but memory was off or never saved it.
4. You expected custom instructions to apply, but they were never set — or you are in a context where they do not apply.

Knowing which one you are in saves you from re-pasting your life story into every chat.

The four kinds of memory

OpenAI documents each of these separately. If you only remember one sentence from this page: a new chat is a new chat. The thread does not follow you.

1. The thread has a working limit

A conversation can only hold so much before the earliest turns stop influencing the answer. When that happens, the model does not announce it. You notice because the answer starts repeating a question you already answered, or forgets a constraint you set at the top.

Signs you are there:

• It re-asks for information you gave in the first message.
• It violates a rule you set earlier (“no bullet points”) without noticing.
• The replies get shorter and more generic.
• You are scrolling a long way to find the original brief.

The fix is not to keep going. It is to re-brief. Write a short handover message:

Before we continue, here is the state of this task:
- Goal: [one sentence]
- Decisions already made: [3-5 bullets]
- Constraints that still apply: [rules, tone, format]
- What is still open: [the next decision]

Do not repeat the earlier discussion. Start from this state and answer the open question.

That message costs 60 seconds and buys another long stretch. If the thread is already a mess, start fresh and paste the handover as the first message. A clean thread with a good brief beats a long thread with a bad one.

2. Saved memory is a preference, not a notebook

When memory is on, ChatGPT can remember useful facts you have shared — your name, preferences, goals — and use them in later conversations. You can view, edit, or delete memories at any time in Settings → Personalization → Memory.

Two things people get wrong:

• Memory is not a document store. It holds facts, not your 40-page brand guide. For documents, use a project or paste the file.
• Memory is not available in custom GPTs. If you are inside a custom GPT and wonder why your saved preferences vanished, that is why.

If you want something remembered, say so plainly: Remember that I prefer British spelling and no exclamation marks. Then check that it actually appears in the memory list. If it is not there, it will not be used.

If you do not want it remembered, turn memory off for that chat or use Temporary Chat.

3. Custom instructions are the standing brief

Custom instructions are the things you would otherwise paste at the top of every chat: your role, your audience, your formatting preferences, the phrases you hate.

Set them at Settings → Personalization on the web, or Settings → Customize ChatGPT on mobile. OpenAI says they apply immediately to all chats, on all plans, across web, desktop, iOS, and Android.

Good custom instructions are short and testable:

I am a [role] writing for [audience].
Default to plain English, short sentences, no exclamation marks.
When I ask for a draft, give me the draft first and the notes after.
If a request is ambiguous, ask one clarifying question before answering.

Bad custom instructions are a mood board: “be creative, be helpful, be human.” They do not change an answer you can point at.

One caution: custom instructions apply broadly, including to tasks where they are wrong. If you write “always answer in Spanish” because you are learning Spanish, every work email draft will also be in Spanish.

4. Projects hold a job, not a conversation

Projects keep related chats, files, and instructions together, and ChatGPT can use that context to support ongoing work. Availability depends on your plan and workspace settings.

Use a project when the work has a durable subject: a client, a book, a course, a codebase. Use a plain chat when the work is a one-off. The failure mode with projects is treating them as a junk drawer — every file, every chat, no structure. A project with three clear files and one instructions block works better than one with thirty uploads.

Privacy: what you paste is data

This is not a memory feature, but it belongs in the same conversation.

• Treat pasted text as data you have handed to a third party. Do not paste passwords, API keys, bank details, medical records, or someone else’s personal data.
• Temporary Chat does not appear in history, does not use or create memories, and is not used to train models. It is the right mode for a sensitive one-off. It is not a guarantee of anonymity; it is a different retention path.
• You can export your data from Settings → Data Controls → Export, and delete your account from Settings or the privacy portal.
• Retention and training rules are set out in OpenAI’s policies. We link them rather than summarise them, because a summary that is six months stale is worse than no summary.

A five-minute setup that pays for itself

Do this once, then stop thinking about memory.

1. Write custom instructions with your role, audience, and three formatting rules.
2. Add only facts that stay true to memory: name, time zone, spelling preference. Not project details, which change.
3. Create one project for your main ongoing body of work. Put the key files and one instructions block in it.
4. Decide your sensitive-data rule now. Ours: no credentials, no client personal data, no medical details, ever.
5. Use the handover message when a thread gets long. Do not let a thread die of old age.

Related

• A prompt is a brief, not a wish — how to write the handover message properly.
• ChatGPT, minus the mystery — the first-session basics.
• The office workflows worth automating first — where projects earn their keep.
• Open the workbench — start a task with a structured brief instead of a fragment.