Graphiti is an innovative framework open-sourced by the Zep team, specifically designed for building and querying time-aware knowledge graphs, optimized for AI agents operating in dynamic environments. This project addresses the limitations of traditional Retrieval-Augmented Generation (RAG) methods in handling dynamic data, providing AI applications with powerful long-term memory capabilities.
Graphiti tracks changes in facts and relationships over time, supporting point-in-time queries. Graph edges contain temporal metadata to record relationship changes, enabling the system to understand the historical evolution of information.
Graphiti incrementally processes incoming data through a real-time, time-aware knowledge graph engine, instantly updating entities, relationships, and communities without batch recomputation.
Provides semantic search, BM25, and graph-based search, with result fusion capabilities to ensure the accuracy and relevance of retrieval results.
Graphiti continuously integrates user interactions, structured, and unstructured data to build a comprehensive knowledge representation.
Graphiti is built on the Neo4j graph database, leveraging the capabilities of LLMs (Large Language Models) to automatically extract entities, relationships, and temporal information. The system is capable of:
A key differentiating feature of Graphiti is its ability to manage dynamic information updates through temporal extraction and edge invalidation processes. The system will:
Supports assistants that learn from user interactions, integrating personal knowledge with dynamic data from business systems such as CRM and billing platforms.
Suitable for various fields such as sales, customer service, healthcare, and finance, providing assistants and agents with long-term recall and state-based reasoning capabilities.
Supports agents that autonomously execute complex tasks, capable of reasoning based on state changes from multiple dynamic sources.
Zep outperformed MemGPT, the current state-of-the-art system, in the Deep Memory Retrieval (DMR) benchmark, demonstrating its superior performance in memory management.
The project provides a new MCP server, providing powerful knowledge graph-based memory capabilities for Claude, Cursor, and other MCP clients.
Although Graphiti was initially developed for Zep Memory, the team realized its potential extends far beyond memory applications and decided to open source it, hoping the community can explore more possibilities.
Graphiti represents a significant breakthrough in the field of AI memory management. It not only addresses the limitations of traditional methods but also provides a technical foundation for building truly intelligent AI systems with long-term memory capabilities. By capturing temporal changes and supporting advanced search technologies, it addresses the challenges faced by existing systems, enabling AI applications to maintain dynamic and accurate information states.