User interest modeling is the key technology of personalized information retrieval. The access behavior, which user browse, save, collect, or print Web page, can really represent user interest. This paper expatiates on the relation between user interest and access behavior on the basis of vector space model and document vectorization, discusses the quantificational formula of user interest degree which is affected by document structure or time forgetting factor, and introduces user interest modeling algorithm. The system structure, result filtration, user interest model maintenance of the personalized information retrieval system based on multi-agent are presented in detail. The experimental result indicates that the system can truly reflect user interest, effectively improve the precision rate and recall ratio of information retrieval, and obviously increase user's satisfaction degree.