The problem of mechanically retrieving references to documents, as a first step to fulfilling the information need of a researcher, is tackled through the design of an interactive computer program. A view of reference retrieval is presented which embraces the browsing activity. In fact, browsing is considered important and regarded as ubiquitous. Thus, for successful retrieval (in many circumstances), a device which permits conversation is needed. Approaches to automatic (delegated) retrieval are surveyed, as are on-line systems which support interaction. This type of interaction usually consists of iteration, under the user's control, in the query formulation process. A program has been constructed to try out another approach to man-machine dialogue in this field. The machine builds a model of the user's interest, and chooses references for display according to its current state. The model is expressed in terms of the program's knowledge of the literature of the field, namely a network of references and associated subject descriptors, authors and any other entity of potential interest. The user need not formulate a query the model varies as a consequence of his reactions to references shown to him. The model can be regarded as a binary classification induced by the user's messages. The program has been used experimentally with a small collection of references and the structured vocabulary from the kedlars system. A brief account of the program design methodology is also given.