Why ChatGPT Keeps Forgetting Context (and 5 Fixes)
You told it the constraints. It agreed. Forty messages later it suggests the exact thing you ruled out.
Or you open a fresh chat and it has no idea who you are, even though you have been using it every day for a year.
Those look like the same problem. They are not, and that is why generic advice does not work. ChatGPT has five separate ways of losing your context, and the fix for one does nothing for the others. This guide helps you tell them apart in about thirty seconds, then fixes each one.
First, ask ChatGPT what it actually used
Before changing any setting, use the diagnostic that is already in the product.
When ChatGPT uses your memory to answer something, a Sources button appears in the row of small icons under that response. Click it and a panel opens listing exactly what shaped the answer, including each memory it pulled in. You can also reach the same thing from the … menu under any response, as View sources.
The absence of that button is the useful part. Ask something ChatGPT should already know about you. If the answer comes back with no Sources button, it used no memory, and your problem is a setting rather than the content of your memory. If the button is there but the memories listed are wrong or stale, the setting is fine and the memory itself needs fixing.
That is a thirty second test, and it tells you which half of this article to read.
One caveat worth knowing: sources are not shown in chats you share with other people, and they may not list every factor that shaped a response.
Now match what you are seeing against the five causes below.
Cause 1: the thread outgrew the context window
What it looks like: it forgets something from earlier in this same conversation. This is the one people usually mean when they say ChatGPT loses context.
Why it happens: the model does not remember your chat the way you do. On every single turn it re-reads the conversation from the top, or as much of it as fits in a fixed budget called the context window. When the thread grows past that budget, the earliest messages stop being read. They are still on your screen. They are no longer in front of the model.
There is a softer version of this that starts earlier. Well before the hard limit, details buried in the middle of a long thread get attended to less reliably than details near the start or the end. A constraint you set in message seven does not vanish, it fades.
The fix: stop trying to save the thread. Start a new one and carry the essentials over.
- In the old thread, ask: summarize the decisions and constraints we have agreed on, as a bulleted list
- Read the list and correct anything wrong. It will miss things
- Open a new chat and paste it in as your first message
The point is that a summary you approved is worth more than forty messages the model can no longer see.
When it will keep happening: any time you work on one thing for hours in a single thread. If that is your normal pattern, put the standing facts somewhere that does not live inside the thread. Custom instructions, a project, or a memory layer that sits outside ChatGPT.
Cause 2: memory is off, or you are in a Temporary Chat
What it looks like: a brand new chat knows nothing about you. Not a fading, a blank.
Why it happens: two possibilities, and they look identical from the outside.
Memory may simply be off. The control is Enable memory, and it lives in Settings, Personalization, near the bottom of that page under the heading Memory. Worth knowing before you go looking: Memory is not its own tab in the settings sidebar, it is a section you have to scroll down to.
Or you are in a Temporary Chat. Temporary Chats do not read your existing memories and do not create new ones. Nothing you say in one will be there tomorrow. This is working as designed, and it is easy to leave one on without noticing.
The fix: check both, in that order. Settings first, then look at whether the chat you are in is temporary.
Cause 3: the memory page does not show everything
What it looks like: you open the memory page, scan it, and the thing you specifically asked it to remember is not listed. So you conclude it did not save.
Why it happens: the memory summary is a high-level view, not a complete inventory. OpenAI's own documentation is direct about this: the summary should capture the most important details, but it will not include everything ChatGPT remembers from your chats.
So an empty-looking memory page is weak evidence. It does not prove ChatGPT forgot.
The fix: ask it directly. In any chat, type what do you remember about my work or what do you know about the project I am building. That queries the actual memory, not the summary view.
If the summary looks stale rather than incomplete, you can force it to rebuild. Go to Settings, Personalization, scroll to Memory, click Manage next to Memory summary, then use the three-dot menu and choose Refresh.
To correct something rather than delete it: type the correction into the text box at the bottom of the memory summary, or highlight the specific line you want changed. You do not have to wipe everything to fix one wrong fact.
Cause 4: you are inside a Project
What it looks like: ChatGPT knows you everywhere except in one workspace. Same account, same settings, and inside that one project it acts like a stranger.
Why it happens: every project is created with one of two memory modes, and the choice is made in a small dropdown in the create dialog that most people click straight past.
- Default memory. The project can use memory from your other chats, and your other chats can use what it learns. This is the option that is pre-selected.
- Project-only memory. The project can only see its own memory, and its memory is hidden from every chat outside it. ChatGPT also notes that Work mode is unavailable for projects set this way.
If a project acts like it has never met you, it was created as project-only. Either you picked that and forgot, or someone else created the project and picked it for you.
Sharing does this too. Once a project is shared with anyone, project-only memory turns on automatically, and shared projects have no access to any member's memories from outside the project.
The fix, and the reason this one is worth reading twice: there is no fix after the fact.
Directly under those two options, in small grey text, ChatGPT says: this setting can't be changed later. It is not a preference you can toggle once the project exists. Removing every collaborator from a shared project does not undo it either.
So the only real fix is to create a new project with Default memory selected and move your chats into it. Which means the useful version of this advice is preventative: look at that dropdown before you click Create project, because it is the one memory decision in ChatGPT you get to make exactly once.
Cause 5: the thing you deleted came back
What it looks like: you deleted a memory. Days later ChatGPT brings it up again.
Why it happens: the memory and the source of the memory are two different objects, and deleting one does not delete the other.
Saved memories are stored separately from your chat history. Delete a chat, and any memory formed from it survives. Turn memory off and back on, and ChatGPT may build new memories out of the chats that are still in your history, including old ones.
The fix: to actually remove a fact, delete every place it lives. That means past chats, archived chats, uploaded files, the memory summary itself, and any connected apps that hold it. One of those alone is not enough.
Two more things worth knowing. Deleting a memory does not edit your old conversations, so a chat where ChatGPT said "since you live in San Francisco" will still say that. And OpenAI retains a log of deleted saved memories for up to 30 days for safety and debugging.
The two memory screens, and which one you have
Halfway through 2026 the memory interface changed, and both versions are still in use. If a guide you are reading does not match your screen, this is usually why.
Both live in the same place, and this is the part that trips people up. Open Settings, Personalization, then scroll past Base style, Characteristics, Custom instructions, and About you. Memory is near the bottom, below all of that. It is not a tab of its own.
Inside it you get an Enable memory toggle and a Memory summary row with a Manage button. The link to the legacy list is not a button, it is the words saved memories inside the small grey description text under Memory summary. Click those two words and you are in the old list. From there, Try improved memory brings you back.
The one thing none of these fix
Every fix above keeps your context inside ChatGPT. None of them move it.
Open Claude and it does not know what you told ChatGPT this morning. Open Gemini and it starts from nothing. If you use one assistant, that is irrelevant. If you use several, it is the version of this problem you hit most often, and no ChatGPT setting addresses it, because the memory belongs to the platform rather than to you.
There are three practical ways to handle it, and they are worth comparing honestly.
Keep a document. A running notes file with your standing context, pasted into the first message of important chats. Free, works everywhere, and you will stop doing it within two weeks.
Export and re-import. ChatGPT can export your data, and some assistants can import memory from other providers. Fine for a one-time move, too slow to be a habit.
Put the memory outside the assistants. Keep one context layer of your own and let each assistant read from it. This is what Unibase Memory does. It captures context from your chats and lets you send the relevant parts into ChatGPT, Claude, Gemini, Grok, Kimi, or DeepSeek, so the same background follows you instead of being retyped.
It is a Chrome extension, free to start, and it does nothing for Causes 1 through 5 above. It solves a different problem: the one where the memory works fine and simply cannot leave.
If you want the manual version first, we wrote out the step-by-step process for moving context between assistants, including what each platform lets you export. And if Gemini is the one giving you trouble, the causes there are different enough to need their own guide.