Why do question-answer pairs that are not in the training data appear in the historical messages of calling the large model?

[Slun12]

After deleting the data in the training set, when asking the same question again, the deleted question and answer pairs still appear in the historical messages. Why is this?