MemoRAG is an RAG framework that leverages a memory model to achieve global understanding of entire databases. It recalls query-specific clues from memory to enhance evidence retrieval and generate more accurate, contextually rich responses. The system includes training scripts, a Lite mode for processing millions of tokens, and was published at theWebConf 2025.