Summary: The problem of memory overflow and the efficient storage and retrieval of temporal knowledge is discussed. For temporal systems, dealing with continuous streams of time-ordered data, the problem of freeing memory is an important one. For humans the problem is one of selective forgetting and reorganisation of memory. There are several approaches to temporal representations which model human mechanisms, e.g. \textit{J. F. Allen}'s interval calculus [Commun. ACM 26, 832-843 (1983; Zbl 0519.68079)], where the mechanisms for information loss are implicit in the relativistic nature of the representation itself or are treated as secondary to the representation. We take the view that the mechanism for information loss is of prime importance and that it can define the representation. The objective is to make the mechanism explicit, by means of a production system architecture. A rule base defines an intelligent forgetting scheme, which manages a database of temporal facts. Two examples of rule bases are given. In the first, it is assumed that restructuring is directed at a particular application, so that `irrelevant' data may be descarded. This assumption is shown to lead to a state-based temporal representation appropriate to the example. In the second, general information loss is modelled, where the resulting data stores are universally applicable, even if incomplete. The starting point for the discussion is a model of temporal data as a matrix of values, temporal order being represented as row order. This background model is used as a universe in which to examine derivable representations. Representations involving data compression without loss of information are examined first. Then principles governing controlled information loss are examined. Information loss is characterised by means of invariance groups of transformations on the data matrix and a model of information loss as a taxonomy of production rules for data deletion is proposed.