Graphiti
used in
0
stacks
Open-source temporal knowledge-graph framework from Zep for building agent memory.
Graphiti is an open-source framework, maintained by Zep, that builds temporally-aware knowledge graphs for AI agents. It ingests data as discrete episodes and tracks how facts and relationships change over time, giving each fact a validity window and full provenance back to source data for point-in-time queries. It supports hybrid retrieval that fuses semantic embeddings, BM25 keyword search, and graph traversal, with custom entity/edge types via Pydantic models. It runs on graph backends such as Neo4j, FalkorDB, and Amazon Neptune, and is the engine powering Zep's managed agent-memory platform.
Strengths
- Bi-temporal model tracks when facts became true and when they changed
- Hybrid retrieval across semantic, keyword, and graph traversal
- Incremental episodic updates without full graph recomputation
- Open source with a managed Zep platform for enterprise scale
Tradeoffs
- Requires running a supported graph database backend
- Graph-based memory adds modeling complexity over simple vector stores
- LLM calls for entity extraction add ingestion cost and latency