getzep/Graphiti
Build Real-Time Knowledge Graphs for AI Agents
Graphiti is an open-source temporal knowledge graph framework by Zep (getzep) that gives AI agents real-time, evolving memory grounded in source data. Instead of flat document chunks, your agent queries a living graph that tracks what is true now and what was true before.
Our Review
Graphiti is one of the most-starred agent-memory projects on GitHub, with 26,000+ stars and a peer-reviewed paper (arXiv:2501.13956) behind its design. Built by the Zep team and released under Apache-2.0, it stores facts as a graph where every relationship carries a start and end date, so retrieval can reason about change instead of plain similarity.
What Graphiti does:
- •Temporal context graphs every fact carries a validity window, so agents know what is true now and what was true before.
- •Hybrid retrieval query across meaning, keywords, and graph relationships in a single call instead of plain vector search.
- •Incremental updates new interactions and documents fold into the graph without recomputing the whole structure.
- •Custom ontology define your own entity and edge types with Pydantic models for domain-specific graphs.
- •Provenance to source every derived fact traces back to the raw episode that produced it, for auditable memory.
- •Multiple graph backends runs on Neo4j and FalkorDB, with Python as the primary SDK.
Graphiti ecosystem:
- •Zep the managed context-graph platform built on Graphiti, for teams that want production retrieval without self-hosting.
- •MCP server a bundled Model Context Protocol server gives Claude, Cursor, and other MCP clients graph-based memory.
- •Research-backed the underlying approach is published in the Zep agent-memory paper (arXiv:2501.13956).
Getting started:
Install with pip install graphiti-core, point it at a Neo4j or FalkorDB instance, and add data with graphiti.add_episode(). Full docs at help.getzep.com/graphiti.
Limitations:
Graphiti is self-hosted only -- you operate the graph database (Neo4j or FalkorDB) and tune retrieval yourself. It is a Python framework, so there is no visual builder or no-code path. Production-scale governance, sub-200ms retrieval, and multi-tenant management require Zep, the paid managed layer. Knowledge-graph modeling has a learning curve if you are new to graph databases.
Our Verdict
Graphiti is one of the strongest open-source options for giving AI agents durable, time-aware memory in 2026. If your agent needs to remember evolving facts about users, products, or processes rather than only retrieving static documents, a temporal context graph is a better fit than plain vector RAG.
For developers, Graphiti is a Python-first framework you embed directly in your stack. The hybrid retrieval (semantic, keyword, and graph traversal in one query) and Pydantic-based custom ontology give you precise control over how memory is modeled, and the bundled MCP server lets coding agents like Claude and Cursor tap the same graph.
Skip Graphiti if you want a turnkey, fully managed memory layer with SLAs and dashboards; that is what Zep, its commercial sibling, is for. If you only need simple document retrieval, a plain vector RAG pipeline is lighter to operate. For a hosted memory API with less setup, mem0 is a simpler starting point.
Frequently Asked Questions
What is Graphiti?
Graphiti is an open-source Python framework, built by Zep, for creating temporal knowledge graphs that serve as memory for AI agents. It tracks how facts change over time, keeps provenance back to source data, and supports hybrid retrieval across semantic meaning, keywords, and graph relationships, making it purpose-built for agents working with evolving real-world information.
Is Graphiti free and open source?
Yes. Graphiti is released under the Apache-2.0 license and is free to self-host as of 2026. You run it against your own Neo4j or FalkorDB database at no licensing cost. Zep, the managed platform from the same team that provides production-grade retrieval and governance, is the paid commercial offering built on top of Graphiti.
How is Graphiti different from traditional RAG?
Traditional retrieval-augmented generation chunks documents and retrieves them by vector similarity, with no memory of how information changes. Graphiti builds a temporal context graph where every fact has a validity window and traces back to its source episode. Choose plain RAG when you only need static document lookup; choose Graphiti when your agent must reason over facts that evolve over time.
What databases does Graphiti support?
Graphiti stores its context graphs in graph databases, with Neo4j and FalkorDB as the supported backends. You install the framework with pip install graphiti-core, connect it to your database, and ingest data as episodes. Because it is a Python framework rather than a hosted service, you are responsible for provisioning and operating that database yourself.
When should I use Graphiti instead of Zep?
Choose Graphiti when you want a flexible open-source core and are comfortable building and operating the surrounding system yourself. Choose Zep when you want a turnkey, enterprise-grade memory platform with security, sub-200ms retrieval at scale, dashboards, and support included. Graphiti is the engine; Zep is the managed infrastructure built around it for production deployments.
How do I install Graphiti?
Visit the GitHub repository at https://github.com/getzep/graphiti for installation instructions.
What license does Graphiti use?
Graphiti uses the Apache-2.0 license.
What are alternatives to Graphiti?
Explore related tools and alternatives on My AI Guide.
Open source & community-verified
Apache-2.0 licensed: free to use in any project, no strings attached. 26,796 developers have starred this, meaning the community has reviewed and trusted it.
Reviewed by My AI Guide for relevance, quality, and active maintenance before listing.
Topics