Graphiti

Graphiti is Zep's framework for building and querying temporal knowledge graphs for AI agents, grounded in a real research paper (arXiv 2501.13956). Unlike static knowledge graphs or flat RAG chunks, Graphiti's context graphs track how facts change over time — what is true now versus what was true before — maintain provenance back to source data, and support both prescribed and learned ontologies. It continuously integrates user interactions and enterprise data into a queryable graph with incremental updates (no full recomputation) and hybrid retrieval across semantic, keyword, and graph traversal.

It is the engine behind Zep's own memory layer, Apache-2.0 licensed, and at 31,800+ stars one of the most-adopted agent-memory projects in existence. The honest framing: this is infrastructure for teams building stateful agents, not a drop-in chatbot memory — you need a graph database (Neo4j or similar), an LLM for extraction, and the willingness to model your domain as a graph. For agents that must remember not just facts but how facts changed, nothing else in the open-source set does this as deliberately.

1. Install: pip install graphiti-core (Python 3.10+).
2. Stand up a graph database (Neo4j supported; see docs for options) and set your LLM provider key.
3. Define your entity/edge ontology or let Graphiti learn it.
4. Add episodes (interactions, documents) — Graphiti extracts entities and builds the temporal graph.
5. Query across time, meaning, and relationships with hybrid retrieval.