getzep/graphiti

What it solves

Traditional RAG methods often rely on static data summarization and batch processing, which struggle with frequently changing information. Graphiti addresses the need for AI agents to maintain a coherent, evolving understanding of the world by tracking how facts change over time rather than just storing static document chunks.

How it works

Graphiti builds a temporal context graph consisting of entities (nodes), facts/relationships (edges), and episodes (the raw data source). Unlike static graphs, it uses temporal validity windows to track when a fact became true and when it was superseded. It supports incremental data updates, meaning the graph evolves in real-time as new information is ingested, without requiring complete recomputation. It also uses hybrid retrieval, combining semantic embeddings, keyword search, and graph traversal.

Who it’s for

It is designed for developers building interactive, context-aware AI applications and production-grade AI agents that require rich, structured memory and the ability to query historical context.

Highlights

• Temporal Fact Management: Tracks the validity of facts over time, allowing queries for what is true now versus what was true in the past.
• Provenance: Every derived fact traces back to its original raw data (episodes).
• Hybrid Retrieval: Combines semantic, keyword, and graph-based search for high-precision queries.
• Incremental Construction: Integrates new data immediately without batch recomputation.
• Flexible Ontology: Supports both developer-defined schemas (via Pydantic) and learned structures that emerge from data.