Graphiti iconGraphiti

Open-source temporal knowledge graph framework from Zep that tracks how facts change over time for agent memory

31k+ GitHub Stars
Apache 2.0 License
2024 Open Sourced

Overview

Graphiti is an open-source framework for building knowledge graphs that change over time. Most graph RAG systems treat facts as static; Graphiti records when each fact became true and when it stopped being true, so when a user says "I moved to Berlin," the old city is marked invalid rather than deleted. Every fact keeps a link to the episode (message, document or event) it came from. The graph updates incrementally as new data arrives, with no full rebuild.

Retrieval combines embeddings, BM25 keyword search and graph traversal. Graphiti runs on Neo4j, FalkorDB, Kuzu or Amazon Neptune and includes an MCP server. It is built by Zep, a Y Combinator company, and is the engine behind Zep's managed context and memory service.

The Verdict

Who Should Use Graphiti?

Best For

  • Agent memory where user facts and preferences change
  • Domains that need history and provenance (support, CRM, compliance)
  • Incremental updates from streaming conversations
  • Teams that want open source with a managed option (Zep)

Not Ideal For

  • Static document lookup where a vector DB is enough
  • Teams unwilling to run a graph database
  • Very cost-sensitive ingestion (LLM extraction on every episode)

What's Great

  • Temporal facts: knows what was true when
  • Provenance back to the source episode
  • Hybrid semantic, keyword and graph retrieval
  • Choice of graph backends
  • Large, fast-growing open-source community

Watch Out For

  • Requires running and tuning a graph database
  • LLM-based extraction adds cost and latency to ingestion
  • Enterprise features (SLAs, support) come via paid Zep Cloud
  • More moving parts than a vector-only memory

Pricing

View all features & details

Key Features

  • Bi-temporal fact validity windows
  • Episodes with provenance
  • Incremental graph construction
  • Hybrid retrieval: embeddings, BM25, graph
  • Custom entity types and ontologies
  • MCP server and REST service

Backends and Integrations

  • Neo4j, FalkorDB, Kuzu, Amazon Neptune
  • OpenAI, Anthropic, Gemini and other LLMs
  • LangGraph and LangChain
  • Claude and Cursor via MCP

How It Compares

Feature Graphiti Cognee Mem0 Microsoft GraphRAG
Temporal facts Yes Enterprise tier No No
Incremental updates Yes Yes Yes Batch rebuild
Open source Apache 2.0 Apache 2.0 Apache 2.0 MIT
Managed option Zep Cloud Cognee Cloud Mem0 Platform No
MCP server Yes Yes Yes No

Is this your tool? Add a badge