How to make ChatGPT remember your conversations

You explained your job, your constraints, what you do not eat. New chat, and it starts from scratch. That is usually not a bug: most of the time it is a setting, and sometimes it is a real limit of the product. The two cases call for different fixes.

Why ChatGPT forgets

A language model retains nothing on its own. It sees only what is in the current conversation — the context window — and that window empties when you open a new thread. Any continuity you notice comes from a layer bolted on top of the model: a feature that writes certain things down and feeds them back later.

In ChatGPT that layer is simply called Memory. According to OpenAI's help page, it can draw on several sources: past chats, saved memories, custom instructions, files in Library and connected apps — depending on your plan and region. If it is off, or not yet available on your account, ChatGPT has no reason to remember anything.

So there are three ordinary reasons it "forgets": the feature is off, it is on but did not judge the detail relevant, or you are in a temporary chat, which neither creates nor uses memories.

Turning memory on and checking it

The path, as OpenAI documents it:

1. Open Settings.
2. Select Personalization.
3. Select Memory.

Depending on your account you will find a general switch and, sometimes, two separate controls: Reference saved memories and Reference chat history. The first covers what you explicitly asked it to keep; the second lets ChatGPT draw on your past conversations. Turning the first off also turns off the second. All of this is in the Memory FAQ.

Once it is on, the simplest way to learn what it kept is to ask. You can also tell it directly to remember something, or to forget it: OpenAI states that these in-conversation requests work alongside the settings.

Review, correct, delete

Two mechanisms coexist, and they do not behave the same way. Saved memories are a list you can read and edit. The memory summary, in the newer experience, is an overview you can correct by editing the text, or wipe in one go with "Delete and turn off memory".

One detail that catches everyone, stated plainly by OpenAI: deleting a chat does not delete the memory that came from it. They are stored separately. To make a piece of information go away you remove the memory *and* the original chat. Conversely, turning off chat-history referencing schedules the derived information for deletion within 30 days, but leaves the chats in place.

On "memory full": in the older list-based experience you could hit a cap and have to prune. The current help page says there is no separate storage limit for what ChatGPT references through chat history. If you still see a full notice, you are on the legacy saved-memories experience — the same page explains how to switch to improved memory.

What this memory will not do

It works well for what it is designed for: making ChatGPT more relevant to you, inside ChatGPT. Its limits are structural, not defects:

• It is tied to one account and one product. What ChatGPT knows about you does not exist for Claude, Gemini, Mistral or the assistant in your code editor.
• You do not choose what is kept. OpenAI says so: memory does not retain every detail; ChatGPT decides what is relevant.
• The summary is not exhaustive. It may omit information derived from sources that are not shown individually.
• It does not export as a memory. What you can retrieve is an archive of conversations, not a set of memories ready to plug in elsewhere.

On that last point, OpenAI documents how to request a copy of your data: it is an account export, not a memory transfer to another product.

When it is more than enough

If you use a single assistant, and it is ChatGPT, built-in memory is the right choice. It is free, it improves with use, and there is nothing to install or maintain. Since June 2025 OpenAI offers it to free accounts too, in a lighter form than for subscribers. Turn it on, check now and then what it has kept, and you are done.

Adding an external tool in that situation is complexity for nothing. We would rather write that here than have you find out later.

When it stops being enough

The problem shows up the day you open a second assistant. Claude for long writing, ChatGPT for the rest, a third one in your code editor: each has its own memory, none of them talk, and you end up explaining the same thing in three places. That is not a flaw in ChatGPT — its memory was never meant to leave ChatGPT.

The fix is to turn the direction around: keep your memories *outside* the assistants, somewhere each of them can read. Technically that means a memory server connected to every assistant through a shared protocol, MCP. You say something once, all your assistants know it. We explain how it works in a separate article on shared memory.

Pryzm Memory is a server of that kind: hosted, connected in two minutes with nothing to install, shared by every assistant you attach. It is only worth it if you use more than one — for a single assistant, reread the previous section.