What the research found
In May 2025, researchers published a study called LLMs Get Lost In Multi-Turn Conversation. In simulations across six generation tasks, where instructions were revealed gradually, the abstract reports an average performance drop of 39% compared with single-turn instructions. In their words, when the models "take a wrong turn in a conversation, they get lost and do not recover".
The study tested 15 models. Section 6.2 reports an average increase in unreliability of 112% in its underspecified multi-turn setting. The researchers suggest consolidating the requirements before trying again. These are results from those simulations, not a prediction of how often your own chat will be wrong.
Why the start of a chat fades
Your early instructions are still in the chat. But by message eighty, they sit alongside drafts, corrections, side questions and ideas you dropped. The model has to weigh all of it, and the rules you set at the start can lose out to whatever came last.
What helps, and what doesn't
- Repeating the rule works for a reply or two, then it fades again.
- Correcting it in the same chat adds one more message to a chat that already has too many.
- Starting a new chat with a short brief puts your rules and decisions back at the top, where they count.
The brief needs four things: the goal, what's decided, the rules you kept repeating, and where you stopped.
Start fresh in one click
Uncook's Handoff writes that brief for you. On the ChatGPT page, click the Uncook icon and pick Handoff. You get a prompt with your goal, what's decided, the rules you set and where you stopped.
Read it, change what you want, then copy it into a new chat.
Questions people ask
Does ChatGPT forget everything in a long chat? No. The earlier messages are still there. They just carry less weight as the chat grows, so the model can drift from them.
Will a new chat never make mistakes? No AI is mistake-free. A new chat with a clear brief simply starts from what you decided, not from a long, tangled history.
Does the same thing happen in Gemini, Claude or Copilot? The study tested models from several providers. It does not establish the performance of every current ChatGPT, Gemini, Claude or Copilot model. A handoff can be pasted into the AI you choose.
Start fresh,
without starting over.
Add to Chrome
We don't keep the conversations you bring. When you click Send to Uncook in the extension window, the extension sends the chat from your tab to our servers, encrypted, for up to 30 minutes, and we delete it as soon as your Uncook page opens it. What Uncook makes from it is kept as a task, for as long as our Privacy page says. Neither your chats nor your tasks are used to train AI models.