This paper presents an approach to design a multi-agent system managing a corporate memory in the form of a distributed semantic web and describes the resulting architecture. The system was designed during the CoMMA European project (Corporate Memory Management through Agents) and aims at helping users in the management of a corporate memory, facilitating the creation, dissemination, transmission and reuse of knowledge in an organisation. The implementation integrated several emerging technologies: multi-agents system technology (using the JADE FIPA-compliant platform), knowledge modelling and XML technology for information retrieval (using the CORESE semantic search engine) and machine learning techniques. Here, we describe the agent roles and interactions, we explain the design rationale for the agent societies and we discuss the configuration and implementation issues.