Digital television and the emerging communication technologies have created an overabundance of programs and information available from which each consumer can choose. The consumer will need new solutions enabling smart and active decision making over viewing preferences. The TV-Anytime Forum seeks to develop specifications to enable audio-visual and other services based on mass-market high volume digital storage in consumer platforms. The central element of TV-Anytime is Personal Digital Recorder (PDR), which is a kind of personal digital media storage device that is widely expected to become an extremely popular consumer electronic device for DTV in the near future. Given the exponential increase on content available from service provider (broadcaster, Internet, etc), software agents have been given a lot of attention lately. Agents have the distinguishing ability to automate repetitive and time consuming tasks, including learning the user’s preference, filtering and recommending content in the PDR. This thesis proposes a multi-agent system for adaptive and personalized services in the TV-Anytime environment. The system is designed to assist users by adapting to their personal preferences. The architecture of the multi-agent system was presented. And the details of these agents, that is, filtering agent, recommendation agent, profiling agent, interface agent, storage management agent, report agent, and privacy protection agent are described. The main contribution of this thesis is to present the architecture of the multi-agent system. To prove the contributions made in this thesis and to implement the system, we have analyzed and designed several key technical points of the system such as: XML-based metadata representation, VSM (Vector Space Model) based feature representation and similarity measurement method, preference knowledge learning and update methods, and KQML-based inter-agent communication. The metadata representation is defined and implemented in XML, which complies with TV-Anytime specifications in general. The vector space model for content feature representation has been generalized for use. The distance metric and cosine correlation for computation of content-profile similarity (content score) have also been generalized. This thesis adopts relevance feedback as the preference knowledge learning method, and instantiates the feedback parameters. This thesis also describes the design of KQML based inter-agent communication and its implementation in constructing the message. The prototype system has been implemented on PC/Linux. We have made experiments and performance evaluation on the prototype. The experiment results are encouraging, which shows that the system proposed in this thesis is useful to consumers. The prototype has demonstrated that our system can really provide personalized content for users.