Why ChatGPT forgets a file you already uploaded

The attachment is still sitting there in the scrollback, which is why this reads like a bug. It is not. That chip is a record of what you sent, not a guarantee the contents are still in front of the model.

Nothing announces the moment it happens either. Answers get vaguer, then generic, then at some point ChatGPT tells you it has no file and asks you to upload one.

What a context window actually is

For every single reply, the model has to hold one bundle in view at once: the system instructions, every message you have sent, every reply it has written, and whatever text was extracted from your file. That bundle has a size ceiling.

While there is room, the file text sits inside it and gets consulted. Once the conversation outgrows the ceiling, something has to leave, and the order is roughly oldest first. Your file arrived at the beginning of the chat, so it leaves at the beginning too.

Rule out the other causes first

Not every case of a forgotten file is eviction.

Test it rather than asking it

When the file drops out, the conversation about the file does not. Earlier summaries, your questions, the points you settled, all of that is still in view and still readable. So it carries on sounding informed while working from its own notes rather than your document, which is exactly where invented quotes and half-remembered figures come from.

Which is why asking whether it still has the file is useless. That question gets a reassuring yes. Ask for something only the file can supply instead:

Quote the first sentence under the heading "Termination", exactly as written.

Three outcomes worth reading carefully. An exact quote means the file is still there. A paraphrase, or a sentence that sounds right but is not in your copy, means it is reconstructing from the conversation. A straight admission that it cannot find the file is the honest failure, and the least common one.

Getting it back mid conversation

1. Ask it to check the file again, once

   ChatGPT can sometimes reopen an attachment rather than relying on what is in the window, so "check the attached file again" occasionally recovers it. OpenAI does not document when this works, so treat it as one attempt, not a reliable fix.

2. Re-attach the file to a new message

   The dependable move. Attaching it again puts the text back in view, along with a short note of what you want from it now.

3. Re-run the quote test

   Before you trust another answer. Ten seconds, and it tells you whether the re-attach actually landed.

Re-attaching into a chat that has run for hours buys less than you would hope, because the window is mostly full of the conversation before the file even arrives. At that point, start again.

Carry the conclusions, not the transcript

When a chat is spent, move your findings rather than your history. Ask for a digest while the file is still present: "List every finding so far as bullets, each with the section it came from." Copy that out, start a new chat, attach the file and paste the digest underneath.

The new chat opens with the document plus a compressed version of the last hour of thinking, in a few hundred words instead of a few thousand. That is a far longer runway than the old chat had left.

Working so it stops happening

One document, one chat. The most effective habit here. Every unrelated tangent is space the file is competing with, and so is every long reply, which is a good reason to ask for bullets rather than essays.

Front-load the extraction. Pull the quotes, figures and structure you need in the first few turns, while the file is certainly present. Analysis of that extract survives eviction. Questions needing a fresh look at page 300 do not.

Send only the part you need. A twelve page section stays in the window far longer than a four hundred page book, and usually produces a sharper answer anyway.

Use a Project for anything you will revisit. Files attached at project level are reachable from every chat inside it, which makes starting fresh cheap. Limits are 25 files per project on Go and Plus, and 40 files per project on Pro, Business, Education and Enterprise. It fixes the new chat case, not eviction inside one long chat.

Custom GPT knowledge suits reference documents. For a style guide or a policy you consult constantly, 10 files per GPT is the allowance, and it covers the lifetime of that GPT rather than each conversation.

The practical decision

If the chat is young, re-attach the file and run the quote test. If the chat is long, take a digest of what you have learned, open a new one, and start again with the file and the digest.

Run that test before you act on anything important. A forgotten file does not produce an error message. It produces a good-sounding answer about a document nobody is reading.