Buildings embody design knowledge that their documentation scatters. Drawings, photographs, specifications, and correspondence are held as separate documents, each one view of the building; the knowledge lives in the fit between the views, and that fit is exactly what generative AI severs twice over: trained on scattered views, its outputs carry no ties to sources. What Buildings Know develops and tests the counter-model: the building-automaton, an AI system that gives one building its own voice. A building's complete records, archival and newly captured, are reconstituted as one multimodal object record with provenance kept per item, and a character layer, defined in four checkable clauses, lets the building answer questions about itself, from its record and only from its record, with every answer carrying its sources. The automaton is composed entirely of open components: public archives, open records, open-weight models running locally, open code. Hosted at TU Wien's Research Unit for Architecture Theory and Philosophy of Technics under Prof. Vera Bühlmann, the project asks what kind of knowledge a building embodies rather than states, and what a digital record must preserve for that knowledge to survive and answer back. It builds automata for two to three canonical Viennese buildings whose archives are public in the host city, measures what the aligned record recovers, and documents encounters with three publics: a design course, a practitioner workshop during a secondment at ATP architekten ingenieure, and a public exhibition, A Question of Character, where visitors converse with the buildings and opt-in dialogues form the first open corpus of conversations between people and buildings. The results: an implementable record schema, an open application, a question map showing what people ask buildings that no archive is asked, three articles, and a documented path by which practices and archives can give their own buildings a voice.