Graphiti: Real-Time Knowledge Graph Framework for AI Agents

Graphiti is a knowledge graph framework designed for AI agents, featuring seamless episodic ingestion of unstructured text and structured JSON while preserving provenance and chronology.
The framework employs a bi-temporal graph model, tracking event occurrence and ingestion times, enabling queries on historical data ("state as of yesterday") and historical truth reconstruction.
Hybrid retrieval at sub-second latency is achieved through a fusion of semantic embeddings, BM25 full-text search, and graph traversals, surfacing relevant edges and nodes.
Custom entity and relationship types can be defined using Pydantic models, allowing for domain-specific schemas and precise knowledge representation.
Scalable and incremental updates are supported through parallelized LLM calls and graph operations, enabling the handling of millions of events without full graph recomputation.
Implementation Detail: Graphiti is based on Neo4j, necessitating consideration of Neo4j licenses. According to reactions, Graphiti is particularly useful once the limits of naive chunk-based RAG are reached, and is a potential game-changer for agentic RAG.