We look at a modification of the traditional vector space model for text representation that allows us to compute vectors for text documents in time. That is, documents are represented by an approximation to their TFIDF vectors without requiring a full corpus of articles. This is achieved by using an exponential window for computing the document frequency of words and by managing a changing lexicon. A graph model is developed in order to provide a reduced representation of the documents. Graph nodes represent document topics and evolve in time. A data set of recent health news articles is used to demonstrate the concepts. Methods are presented for visualizing the data and the content of the graph nodes.