It doesn’t work that way. It ”remembers” that many answers and use that as keeping context and using it as base of new answers.
Propably use are using too many tokens or something similar and it just stops fullfilling your request.
I guess error log should show something.