We propose a system that registers and retrieves text documents to annotate them on-line. The user registers a text document captured from a nearly top view and adds virtual annotations. When the user thereafter captures the document again, the system retrieves and displays the appropriate annotations, in real-time and at the correct location. Registering and deleting documents is done by user interaction. Our approach relies on LLAH, a hashing based method for document image retrieval. At the on-line registering stage, our system extracts keypoints from the input image and stores their descriptors computed from their neighbors. After registration, our system can quickly find the stored document corresponding to an input view by matching keypoints. From the matches, our system estimates the geometrical relationship between the camera and the document for accurately overlaying the annotations. In the experimental results, we show that our system can achieve on-line and real-time performances.