Why does ChatGPT forget your last conversation?

What the model actually sees

Every reply is a fresh request. The model receives a system prompt, whatever memory the product chose to inject, the recent messages in the thread, and your new message. That is the entire input. It does not have a database it can query, and it does not remember yesterday unless something wrote yesterday down.

This is why the same model can feel brilliant in one thread and useless in the next. Nothing changed about the model. The context changed.

Context window is not memory

The context window is short-term working space. It holds the conversation you can see, up to a token limit. Memory is a separate feature that saves a small set of facts, preferences, and instructions, then adds them to future requests. A large context window lets you paste more documents into one conversation. It does not make the app remember you next week.

Why long conversations degrade

When a thread grows past the window, the app has to drop or summarise older messages. Summaries lose detail, and the details are usually the point: the constraint you mentioned in message three, the name you corrected in message nine, the format you asked for. The model is not ignoring you. The information is no longer in front of it.

Four fixes that work today

• Turn memory on and review it. Delete wrong entries; a bad memory is worse than no memory.
• Put stable instructions in custom instructions rather than repeating them in every thread.
• Use projects or folders to keep related chats and files together, so a new thread starts with the right context.
• Keep a short project brief in a document and paste it at the start of a new thread. Three paragraphs of context beat thirty messages of improvisation.

When the fix is a different app

These fixes work inside one vendor. They stop working the moment you switch models or tools, because each product keeps its own memory and none of them export it in a form another app can use. If your work spans several models, the thing you want is memory that sits above them rather than inside one of them.

Swirl is built that way. It keeps a visible memory list that you can edit and delete, and it carries that memory across models and conversations. You can start with a fast model and move to a heavy one without re-explaining the project. Five free messages are enough to see whether the difference is real for your work.

What good memory looks like

• Visible: you can read every stored entry, not just trust that it exists.
• Editable: correcting a wrong fact takes one click.
• Portable: it survives switching models and starting a new thread.
• Scoped: project memory stays with the project instead of leaking everywhere.
• Deletable: removing an entry actually removes it.

Judge any app against those five. Most chat products fail at least two of them, which is why the forgetting problem feels unfixable even when a memory setting is switched on.