The Remembrance Agent (RA) is comprised of two parts, the user interface and the search engine. The user interface continuously watches what the user types and reads, and it sends this information to the search engine. The search engine finds old email, notes files, and on-line documents that are relevant to the user's context. This information is then unobtrusively displayed in the user's field of view.
Currently, the user interface runs in elisp under Emacs-19, a UNIX based text editor that can also be used for applications such as email, netnews, and web access. The Remembrance Agent displays one-line suggestions at the bottom of the emacs display buffer, along with a numeric rating indicating the relevancy of the document. These items contain just enough information to represent the contents of the full document being suggested. For example, the suggestion line for a piece of email includes the subject line, the sender, and the date received. The suggestion line for a notes file contains the file name, owner, date last modified, and the first few words of the file. With a simple key combination, the user can display the full text of a suggested document. The RA may suggest files based on the most recent 500 words of the user's current document, the most recent 50 words, the most recent 10 words, or it may display suggestions in a hierarchy of all 3 contexts. In this way, the RA can be relevant both to the user's overarching context and to any immediate associations. Each user may customize the number of suggestions displayed, the type of documents referenced, and the update frequency of the display.
The current implementation uses Savant, an in-house information retrieval system, which is similar to the SMART information retrieval program [Salton, 1971]. This search engine determines document similarity based on the frequency of words common to both the query and reference documents. In order to improve its speed, Savant creates an index of all words in each document source nightly. While Savant is not as sophisticated as many more modern information-retrieval systems, it does not require human pre-processing of the documents to be indexed.
Figure:
An example of Remembrance Agent output while editing an early
version of this document. The first number on each line provides a
file label for convenience, while the second number is the ``relevance
measure'' of the message.
Figure 3 shows the output of the Remembrance Agent (in the bottom buffer) while editing an earlier version of this document (top buffer). The reference database for the RA was the third author's e-mail archives. Personal email and notes files are good sources for an RA to reference, because these files are already personalized and automatically change with the user's interests. This personal reference database is much richer in potentially relevant information than, for example, the Web. Such personalization also helps solve many ``homonym'' problems. For example, when an AI researcher mentions ``agents'' in a memo or paper, that person's RA will suggest work on autonomous agents. A chemist's RA, on the other hand, will bring up lists of various solvents and other chemical agents. Additionally, because the reference database is the user's own work, the one-line summary may be all that is necessary to remind the user of a document's relevancy. Finally, given the greater opportunity to record personal and experiential knowledge using a wearable computer, the user generates a fertile reference database for the RA.
While the Remembrance Agent gains most of its contextual information from typed text, wearable computers have the potential to provide a wealth of contextual features similar to those discussed in [Lamming and Flynn, 1994]. With the systems described in later sections, additional sources of user context information may include time of day, location, biosignals, face recognition, and visual object tags. Thus, the RA begins to show the advantages of wearable, context-driven augmented reality. However, with a more comprehensive agent, the wearable computer may be able to uncover trends in the user's everyday life, predict the user's needs, and pre-emptively gather resources for upcoming tasks.