Chatbots get worse over a long conversation, partly because their own earlier replies sit in the history and pull later answers back toward an early mistake the user already corrected.
A separate model reads only the user's messages, writes down what was asked, checks the assistant's earlier replies against that, then either keeps them with the summary or throws them out and rebuilds the history from the request.
On a test where discarding every assistant reply is known-best, it recovers 55 to 84 percent of the lost ground on three of four tasks and beats that baseline on the fourth. The checker itself gets contaminated by seeing the assistant's reasoning, so it is structurally blinded.