qhjqhj00/MemoRAG

Health v2 · maintenance only

Adoption not part of the score

2266 stars · 159 forks observed · 2026-09-26

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

MemoRAG is a Python RAG framework that uses a super-long memory model to build a global understanding of an entire corpus (up to ~1M tokens) and recalls query-specific clues to improve evidence retrieval and answer generation. It ships with pretrained memory models on HuggingFace, a Lite mode, and training scripts/datasets.

Use cases

• build a RAG pipeline over very large document collections
• answer questions over a million-token corpus with global context
• improve retrieval quality beyond standard vector-search RAG
• run memory-augmented QA over long PDFs or books
• fine-tune a custom memory model for domain-specific RAG
• generate context-rich answers when queries have implicit information needs

When to choose

• your corpus is too large for a single LLM context window
• standard RAG retrieval misses globally relevant evidence
• you want an open-source, research-backed RAG framework with pretrained memory models
• you need to recall query clues from a global memory of the whole database

When to avoid

• you lack GPU resources for running 7B memory models
• your corpus is small and fits in a standard LLM context
• you need a fully managed hosted RAG service
• you only need simple keyword or vector search without generation

Facets

framework · maturity active

rag llm-inference machine-learning search-engine large-language-models artificial-intelligence python cross-platform memory-model long-context knowledge-discovery retrieval huggingface retrieval-augmented-generation natural-language-processing gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="qhjqhj00/MemoRAG")