Temporal reasoning is recognized as a key problem in many AI areas, namely knowledge bases, natural language processing and planning. The ability to deal with partial knowledge is particularly important in a temporal domain. We describe a temporal language that accounts for incompletely specified temporal information about propositions. The language is semantically based on the notion of maximal interval, the denotation of a proposition being a set of maximal intervals where it holds. The main differences between classical formalisms such as those by Allen, McDermott, Shoham and Kowalski and our approach are briefly discussed. In a partial KB, abduction on the temporal order is generally needed to answer a query, and the answer is then conditional on the abduced facts. To comply with the intended semantics, an implicit form of temporal consistency has to be enforced, and this presents the main challenge to the design of the inference mechanism. We present here the syntax and declarative semantics of a propositional version of the language of maximal intervals and a first discussion of the problems in designing an inference system adequate to work with this temporal framework.