The Content-Based Image Retrieval (CBIR) system is a special kind of knowledge-based multimedia retrieval system. Relevance Feedback (RF) is a widely used technique in CBIR for incorporating the user’s knowledge with the learning process, which has been shown to significantly increase the retrieval accuracy. However, the user preferences obtained through RF are often discarded at the end of search, thus requiring the feedback process to restart for each new query. This paper proposes a Long-term knowledge-based Multimedia retrieval System (LMS) based on Latent Semantic Indexing (LSI) and human interaction (RF). Experiments show the effectiveness of the proposed system. © 2007 Inderscience Enterprises Ltd.