getzep/graphiti

A Python framework for building and querying temporal context graphs that track how facts change over time for AI agents.

What it is

Graphiti builds context graphs from structured and unstructured data, where every fact carries a validity window and traces back to the raw episode that produced it. Unlike static knowledge graphs or batch-oriented GraphRAG approaches, it integrates new data incrementally without recomputing the whole graph, and supports historical queries about what was true at a given point in time. It is aimed at developers building context-aware, memory-equipped AI agents that operate on evolving real-world data.

Key features

Bi-temporal fact management: facts have validity windows and are invalidated rather than deleted when information changes, so you can query current or historical state.
Episodes and provenance, with every entity and relationship tracing back to the raw data that produced it.
Prescribed or learned ontology, with developer-defined entity and edge types via Pydantic models.
Incremental graph construction that integrates new data immediately without batch recomputation.
Hybrid retrieval combining semantic embeddings, keyword (BM25), and graph traversal.
Pluggable graph backends and an included MCP server plus a FastAPI REST service.

Tech stack

Primary language: Python, requiring >=3.10,<4; distributed as graphiti-core (version 0.29.2 in the manifest), built with hatchling.
Core dependencies include pydantic>=2.11.5, neo4j>=5.26.0, openai>=1.91.0, tenacity>=9.0.0, numpy, python-dotenv, and posthog.
Graph backends via extras: Neo4j (default), FalkorDB (falkordb>=1.1.2,<2.0.0), embedded FalkorDB Lite (Python 3.12+), Amazon Neptune, and Kuzu (deprecated).
LLM and embedding providers via extras: OpenAI by default, with Anthropic, Groq, Google Gemini, Voyage, and OpenAI-compatible or local endpoints supported.

When to reach for it

Giving agents structured, evolving memory instead of flat document chunks or raw chat history.
Applications that need precise historical queries and automatic handling of changed or contradictory facts.
Continuously ingesting user interactions and enterprise data into a queryable graph in near real time.
Combining semantic, keyword, and graph-traversal retrieval in one low-latency query path.

When not to reach for it

One-shot static document summarization, where a batch GraphRAG approach may be simpler.
Teams unwilling to run and operate a graph database backend and the surrounding retrieval system themselves.
Pipelines relying on smaller or local models that do not reliably honor structured output, which the README warns can cause ingestion failures.
Cases where you want a fully managed, turnkey platform rather than a self-hosted OSS core.

Maturity signal

Graphiti is actively maintained, with a recent last push, an Apache-2.0 license, and a published arXiv paper backing its architecture. It is the open-source engine underneath Zep's commercial context infrastructure, which signals sustained corporate investment. The large open-issue count relative to a project of this size reflects an active, fast-moving codebase and a broad support surface across multiple graph and LLM backends rather than neglect.

Alternatives

Microsoft GraphRAG — use it for static document summarization and batch-oriented entity clustering, per the README's own comparison.
Zep — use the managed platform when you want turnkey, production-ready context retrieval with governance, SLAs, and built-in user and conversation management.
Mem0 — use it when you want a lighter agent-memory layer without operating a full temporal graph engine.

Notes

Kuzu support is deprecated because the upstream project is unmaintained, and the driver emits a DeprecationWarning; new projects should use Neo4j or FalkorDB. Ingestion concurrency defaults low (SEMAPHORE_LIMIT=10) to avoid provider 429 rate-limit errors and can be raised if your provider allows. Graphiti collects anonymous, opt-out telemetry via PostHog, and the README documents exactly what is and is not collected.

Wiki: getzep/graphiti