HistoRAG introduces architectural changes to Retrieval-Augmented Generation based on historiographical principles. It separates retrieval and generation, implements temporal windowing for balanced source representation, and uses LLM-as-judge evaluation for transparent relevance judgments. Evaluated on 102,189 Der Spiegel articles (1950-1979), the framework addresses deficiencies in standard RAG, including temporal skew and weak retrieval correlation, and proposes Zwischentexte as a responsible integration method for LLM-generated content in scholarly work.