getzep/graphiti is a Python project that builds real-time knowledge graphs for AI agents. The latest release is v0.31.1. It unifies facts obtained from conversations and multiple data sources into a graph, and its core is tracking how facts change over time. It is released under the Apache-2.0 license.
If a static knowledge graph only holds facts as they are at present, graphiti aims at a temporal graph that maintains not only how facts changed but also the provenance of the underlying data. It continuously integrates user interactions, structured and unstructured enterprise data, and external information into a queryable graph. On that basis it takes the stance of replacing RAG for finding relevant documents. An ontology is the frame that determines which concepts and relations the graph handles. graphiti supports both predefined and learned ontologies. The README cites the arXiv 2501.13956 paper and explains that it runs linting, unit tests, and type checks in CI.
The repository has 194 files in graphiti_core, 87 files and 121 test paths in tests, 70 files in mcp_server, and 46 files in examples. GitHub stars stand at 31,786. That said, this scale and these CI items alone do not confirm real performance or the benefit of adoption. The given description states it aims to replace RAG, but it presents no comparative performance numbers. So the structure for dealing with facts and their provenance over time is worth noting, but whether it is better than RAG has to be checked separately.