Long chats forget
Every AI model reads your chat through a working memory with limited room, the context window. The longer the chat, the more crowded it gets, and details from the beginning fall out of focus. This is called context rot: you say at message 1 that the place of jurisdiction is Frankfurt, and at message 60 the draft says Munich. It isn't a glitch, it's how the memory works. Studies found it in all the leading models they tested.
Long chats also hold on to mistakes
The opposite problem is context poisoning. If a wrong fact gets into the chat (the AI reads a notice period as 4 weeks instead of 3 months), it stays in the history. Even after you correct it, both the mistake and the correction are read along with every new message, and the mistake can come back in the next draft.
What helps
- A new chat instead of a correction. New topic, or a mistake in the history: start fresh. It's one click, and your old chats and files stay in your workspace.
- Put checked facts first. In the new chat, start with the key points you know are correct.
- Upload instead of paste. The document is the source, not the chat history. Ask for the passage with a page reference.
New chat, same chat or knowledge base?
Rule of thumb: if you upload the same document for the third time, it belongs in a knowledge base.
The indicator in the chat
Next to the message box, a small indicator shows how full the current chat is. It turns from green to yellow, orange and red. When it's orange or red, click it to start a new chat. Long chats also use more of your allowance, because every message sends the whole chat so far along with it. See Models and usage.