MemoRAG is an open-source retrieval-augmented generation framework with 2,200+ stars on GitHub that gives RAG a long-term memory, so systems can answer questions whose evidence is implied rather than stated outright.
Standard RAG works when a query’s information need is explicit, but struggles with ambiguous questions, multi-hop reasoning and tasks that require a global understanding of a whole corpus. MemoRAG, accepted at The Web Conference 2025 and actively developed through 2025, adds a memory-based data interface on top of retrieval: a highly efficient, super-long memory model first reads and remembers the entire database, builds a global understanding of it, and then recalls query-specific clues from memory. Those recalled clues guide evidence retrieval, producing answers that are more accurate and contextually rich than similarity search alone. The project ships the framework, the memory model on Hugging Face, and — since April 2025 — the training scripts and training dataset, so researchers can reproduce and extend the approach rather than treat it as a black box.
MemoRAG is free and open source under the Apache License 2.0. You download the code and models and run them on your own hardware; there is no hosted tier or commercial license.
RAG researchers and engineers building question answering over large, complex corpora — enterprise knowledge bases, research collections and long-document archives where plain vector search keeps missing the point.