In this paper, we have presented WordSieve, a new algorithm for characterizing a user's context by analyzing documents which the user is accessing. The algorithm builds a user profile during document access to reflect the range of user interests, and generates context profiles to reflect the user's current context. WordSieve outperforms TFIDF in initial experiments on associating documents to the contexts in which they were accessed. This performance gain does not seem to be specific to some subset of the overall data, but appears to generalize well over various subsets of the data which we have examined. Research on WordSieve and its application suggests a number of questions for context studies research, especially concerning the use of context in Intelligent Information Agents and the kinds of information about a user's context that can be learned automatically from implicit feedback.