We investigate belief change in the context of man-machine dialogue. We start from Perrault's approach to speech act theory [12], which proposes on default mechanisms for the reconstruction of the agents' beliefs. Then we review Sadek's [16] critiques and introduce his approach. We point out some shortcomings and present a new framework for the reconstruction of beliefs which contrarily to Perrault's and Sadek's is monotonic. We focus on a particular application, viz. cooperative man-machine dialogues. Our basic notion is that of a topic: we suppose that we can associate a set of topics to every agent, speech act and formula. This allows to speak about the competence of an agent, and the preservation of beliefs. We give an axiomatics and a possible worlds semantics, and we show how the belief state of an agent can be reconstructed after a speech act.