FalkorDB is a property graph database aimed at knowledge graphs for large language models, with GraphRAG and agent memory as its main use cases. It represents a graph's adjacency matrix as sparse matrices and executes queries with linear algebra, using GraphBLAS for storage and traversal, with the goal of low query latency.
It runs as a Redis module, so any Redis client can send its graph commands, and it supports the OpenCypher query language with its own extensions. Nodes and relationships carry attributes under the property graph model. Official clients exist for Python, Java, Node.js, Rust, Go, and C#, alongside community clients for Ruby, PHP, Elixir, Julia, and other languages. A browser UI ships with the container image.
Use cases include combining LLMs with domain knowledge graphs to ground responses, pairing graph traversal with vector search for personalized agent applications, persistent conversation memory for chatbots, fraud-ring detection across IPs, devices, and transactions, and multi-tenant security analytics. One instance holds many separate graphs, which serves multi-tenant deployments.
Features
- Sparse matrix storage: adjacency matrices stored as sparse matrices through GraphBLAS
- Linear algebra queries: query execution expressed as linear algebra operations
- Property graph model: nodes and relationships with attributes
- OpenCypher: compatible with the OpenCypher query language, plus proprietary extensions
- Redis protocol: commands callable from any Redis client
- Client libraries: official clients for Python, Java, Node.js, Rust, Go, and C#
- Multi-graph tenancy: many separate graphs (tenants) in one instance
- GraphRAG and agent memory: knowledge graphs combined with vector search for retrieval and long-term memory
