Abstract : This project studies personalized proactive information filtering agents that pushes relevant information to the user without requiring explicit user query. To do this, the agent adaptively learns a detailed user model while observing and interacting with the user. We use Bayesian statistical theory and machine learning techniques to tackle the following two major challenges. We studied two major problems: how to build an initial user profile with minimal user effort, and How to improve personalized recommendation based on multiple evidences, such as social networks, implicit user feedback, and explicit user feedback and context information. This project led to 1 book chapter, 2 journal paper, 4 conference publications and one demo system.