getzep/graphiti

What it solves

Graphiti addresses the limitations of traditional Retrieval-Augmented Generation (RAG) by moving beyond static document chunks. It provides a framework for building temporal knowledge graphs that track how facts, entities, and relationships evolve over time, allowing AI agents to understand not just what is true now, but what was true in the past.

How it works

Graphiti ingests structured and unstructured data into a graph structure where every fact has a "validity window." It uses LLMs to extract entities and relationships from raw data (episodes) and stores them in a graph database (such as Neo4j or FalkorDB). Because it tracks the temporal history of these facts, the system can perform hybrid retrieval—combining semantic, keyword, and graph-traversal searches—to provide context-aware responses without needing to recompute the entire graph.

Who it’s for

This framework is for developers building AI agents or applications that require deep, evolving context, such as long-running conversational assistants, personalized enterprise tools, or systems that must track changing real-world data.

Highlights

• Temporal Fact Management: Automatically tracks when facts become true and when they are superseded, preserving a full historical record.
• Provenance Tracking: Every derived fact is linked back to the original source data (episodes) for full transparency.
• Hybrid Retrieval: Combines semantic embeddings, keyword search, and graph traversal for high-precision, low-latency queries.
• Flexible Ontology: Supports both developer-defined entity types (via Pydantic) and emergent structures learned from data.
• Incremental Updates: Integrates new information in real-time without requiring batch recomputation of the graph.