Why Does ChatGPT Forget Everything? (And How to Stop Re-Explaining Yourself)

You open a new chat and start typing a problem you have been chewing on all week. The answer comes back competent, general, aimed at nobody. So you back up and explain who you are, what you are building, and who it is for.

You have written that paragraph before. Last month, at length, in a window that is still sitting in your sidebar.

There is a technical reason for the gap, and it is not the one most people assume. Your assistant did not forget you. In the strict sense, it never knew you at all.

Why does ChatGPT forget everything between chats?

Because the model has no memory of its own. Text goes in, text comes out, and nothing persists on the far side of that exchange. Every fact an assistant appears to recall about you was quietly re-inserted into the conversation by the product wrapped around the model, a fraction of a second before your message was sent.

That wrapper is doing more work than most people realize, and less than they need. It can carry a handful of saved notes. It cannot carry the thing you actually want carried, which is a real account of who you are, because nobody ever wrote one.

So the question worth asking is not why did it forget. It is what can it see when I hit send.

What the model can see	Is it there?
The message you just typed	Always
The rest of this conversation	Until the window fills, then the earliest parts thin out
Your custom instructions	Only if you wrote them, and only in that tool
Saved memories	Only the notes the assistant chose to keep
Files in a project or workspace	Inside that project, not outside it
Everything else you have ever told it	No

That last row is the whole problem. A year of conversations is not context. It is a pile of transcripts sitting in a sidebar, and no model reads your sidebar before answering.

Memory features help. They do not close the gap.

Turn memory on and things do improve. The assistant stops asking what you do every third session, remembers the format you like, greets you like it has met you. That is genuinely better than nothing, and if you have not switched it on, switch it on.

But look at what those notes actually are. They are fragments a model wrote down mid-conversation because something seemed worth keeping. Prefers bullet points. Is working on a pitch deck. Lives in Atlanta. Individually true, collectively incoherent, and weighted hard toward whatever you happened to discuss in the last few weeks.

Three limits show up fast.

The notes are chosen for you. You did not decide what mattered; a model did, in passing, based on a conversation about something else.

They are recency-heavy. A project you abandoned in March can outrank the work you have been doing every day since, because the model saved one and never got around to saving the other.

And they are locked in one account. Nothing you build inside ChatGPT's memory reaches Claude or Gemini, and there is no export that would let it. Each tool keeps its own private notes in its own private room.

So the familiar feeling is real, and it does not travel. Open a different tool, or a new tool next spring, and you are a stranger again.

The re-explaining tax is not the minutes

It is tempting to file this under inconvenience. Two minutes of setup per chat, a small tax, annoying but survivable.

The real cost is what you produce under that tax. Nobody types a careful account of themselves at the top of a Tuesday afternoon chat. You type the shortest version that gets the model moving, which flattens your situation into something generic — and then the answer comes back generic, because generic is exactly what you fed it.

Then you blame the prompt. You rewrite it, sharpen it, add three more instructions, and the output improves a little without ever becoming yours. That is because the missing piece was direction, not instruction. The model was never short on cleverness. It was short on you.

How to stop starting from zero

The fix is one sentence: write the durable parts of your context down once, in a file you own, and load a copy into every tool you use.

That is it. It sounds too simple to matter, and it is the difference between an assistant that guesses and one that starts from the truth.

The sorting rule is what makes it work. If something will still be true in six months, it goes in the file. If it is a fact about one particular week, it is a note, and notes rot. Your role, your audience, your voice, your constraints, what you are building and why — those stay true. What you had for lunch during a brainstorm does not.

RUMO calls each of those durable pieces a context anchor. Six of them, kept as separate documents so you can update the one thing that changed without rewriting the rest.

You do not need all six to feel the difference. One good file, pasted at the top of a new chat, and the model stops asking what you do for a living.

If the blank page is what stops you, that is normal, and it is a known problem with a name. Context mining is the guided version: answer questions someone else wrote, and the durable stuff comes out of you sideways, which is easier than staring at an empty document trying to describe yourself.

Which one to write first

Start with values, not facts. Your beliefs, your non-negotiables, how you make decisions when they are close. That is the layer everything else stands on, and it is the layer a model can never infer from a transcript.

You can write one free at /anchors/constitution. It takes about half an hour, and it works in every tool the moment it exists.

Then keep it current. A stale file is worse than a thin one, because the model will act on it with complete confidence and no idea it is holding you in a position you left in the spring. Update on real triggers — new role, new client, finished project, changed mind — not on a calendar. Everything smaller than that is ordinary drift, and drift is fixed by pasting the file in again, not by rewriting it.

The next new chat

Nothing here makes an assistant remember you. That is the part worth being honest about: there is no method that gives a model true memory of your life, and anyone selling one is selling a wrapper around the same trick.

What you can change is what the model has in front of it when the window opens. Right now that is one hurried message and whatever notes a vendor happened to save. It could be a page you wrote on purpose, that says who you are, that you can hand to any tool in ten seconds.

So the next time a new chat asks what you do for a living, you do not spend two minutes rebuilding yourself from memory.

You paste.