In this paper, we have presented WordSieve, an algorithm for indexing documents for personal information agents. WordSieve builds a context profile by reading the documents the user accesses and extracting terms which tend to partition the user's browsing behavior into discrete tasks, in order to generate context-sensitive indices that can be used to suggest relevant documents in the future. In our experiments, WordSieve outperforms TFIDF at generating indices that associate documents with explicit task descriptions that were given to the user to guide browsing, but were unavailable to both programs when they generated their indices. We believe that automated context-based indexing techniques such as WordSieve are potentially useful for in many different kinds of environments where personal information agents must assist users in organizing and appropriately retrieving digital resources.