Context window, and why your long chats get worse

A context window is the amount of text a chat can hold in view at one time. Your typed messages, the replies, and any file you paste all sit inside it. Once the conversation outgrows that space, the earliest parts stop being read. The answers get vaguer, and the chat starts contradicting what you settled an hour ago.

What can ChatGPT actually do for me in a long chat?

In a long chat ChatGPT will hold a pricing sheet steady for a while, then it stops holding it. The document shows that before you do. Say you export cotton home textiles and you open a chat to build a distributor pricing sheet. You paste the landed cost workings, set a margin floor, and agree a column order. The first tables come back clean. Much later in the same chat you ask for a revised sheet. The margin floor is gone. A column has been renamed, and a freight figure you corrected early on is back in its wrong form. Nothing broke. The early instructions fell out of the window, and the sheet was rebuilt from whatever was still in view.

Start a fresh chat with a brief, not a scroll

When a chat starts drifting, stop scrolling up to argue with it and open a new one. Paste a brief you keep in a file: your role, the business context, the task, the output format, and the constraints that must hold. Add the few facts the work depends on, like the margin floor and the corrected freight figure. That brief is the part worth saving. Write it once for each piece of recurring work. I keep mine in one file and fix it the day an output comes back wrong. Some people call the slow decay context rot. No bigger tool removes it. A brief you can reload in seconds makes starting again cheap, and the judgment to spot the drift early is what triggers it.

What else should I know about the context window?

Does a longer context window fix this?

A longer window pushes the problem further out, it does not remove it. You still see the same drift in the output, a dropped constraint or a reverted figure, just later in the chat. Treat capacity as headroom, not as a reason to keep one conversation running for a week.

How do I know when to start again?

Start again the moment an output repeats something you already corrected, or drops a constraint you set. That is the signal, not the clock. Copy the brief into a new chat, paste the current version of the document, and carry nothing else across.

Will ChatGPT memory or a Project solve it?

Memory and Projects hold standing facts and files across chats, which helps. Neither keeps a single long conversation accurate. The task brief still has to be restated for each run, and you still check the output against the constraints you wrote down.