A common complaint from ChatGPT users goes like this: the assistant seemed sharp yesterday, and today, in the same conversation, it answers as if it has forgotten everything. A widely repeated explanation blames a rule that uploaded files “expire after three hours.” The truth is more useful than the rumor, and understanding it makes the tool noticeably easier to work with.
There is no three-hour memory timer
ChatGPT does not run a countdown after which the model loses access to a conversation’s content. OpenAI’s help documentation on chat and file retention describes how chats and files are stored and deleted, and none of it describes a fixed interval after which the model stops reading what is in a chat. Regular and archived chats stay in an account until the user deletes them, and the text of a conversation remains available to the model when the chat is reopened.
The three-hour figure most likely comes from a different, narrower behavior. When ChatGPT runs code to analyze an uploaded spreadsheet or document, it does so in a temporary computing environment. That environment is shut down after a period of inactivity, and files that lived only inside it are gone when it closes, which produces the familiar “file is no longer available” message. OpenAI’s documentation notes that files not saved to a user’s Library can expire separately from the chat itself. That is a limit on a sandbox, not on the model’s memory, and it does not affect the conversation text or files saved to the Library or a project.
The real constraint: the context window
What usually causes the “it forgot everything” experience is the context window. Every large language model can attend to only a limited amount of text at once. That budget, measured in tokens (roughly three quarters of a word each), has to hold the system instructions, the conversation so far, and whatever the model has extracted from any uploaded files.
As a chat grows, the oldest material is the first to fall outside that budget. The messages are still visible to you when you scroll up, but the model is no longer reading them when it composes a reply. The result is exactly what users describe: the file is right there in the thread, yet the answers become generic, as if the model never saw it.
This is tied to the length and complexity of the conversation, not to the clock. A short chat left alone for a month will pick up where it left off. A very long chat can lose track of an early file within a single afternoon. It is also not a bug. It is the design of every current chat model, and OpenAI does not claim that ChatGPT will retain the full content of old files indefinitely within a conversation.
What to do about it
A few habits deal with the problem reliably:
- When returning to an old chat, restate the background in two or three sentences before asking a new question.
- If a file matters, upload it again, or keep it in a ChatGPT project or the Library, where it persists independently of any single conversation.
- Start a fresh conversation for a new task rather than appending to a long one; a shorter thread leaves more of the context budget for the work at hand.
- Ask the model directly whether it still has access to the earlier material. It is not a magic command, but the answer tells you what the model is actually working with.
Professional users treat this “context refresh” as routine rather than as a workaround.
The underlying mismatch
People are used to systems with permanent memory: a colleague remembers last week’s meeting, a filing cabinet keeps what was put in it. A chat model works differently. It reasons over whatever is in front of it at the moment of the request, and nothing else. ChatGPT does not become less capable over time, and no timer switches its memory off. Users who internalize that difference tend to get far better, more consistent results from the tool.
