Graphiti

Graphiti is an open-source framework for building and querying temporal context graphs for AI agents. It records when facts become valid and when they are superseded, which supports queries about the present or a past point in time. New structured and unstructured data can be integrated incrementally, and entities and relationships link back to their source episodes. Retrieval combines semantic search, keyword search, and graph traversal. Developers can define entity and edge types with Pydantic models or let structure emerge from data. Listed graph backends include Neo4j, FalkorDB, and Amazon Neptune; Kuzu is marked deprecated. The framework defaults to OpenAI and also supports Anthropic, Google Gemini, and Groq, plus compatible endpoints for OpenAI-compatible providers. It includes an MCP server and a FastAPI REST service. Graphiti is self-hosted and self-managed, so users operate the surrounding system themselves. Requirements include Python 3.10 or higher and an LLM API key; structured-output-capable LLM services work best. The repository lists an Apache-2.0 license. Anonymous usage statistics are collected, with an environment variable available to disable telemetry.

Graphiti is built for temporal, provenance-aware context graphs, with incremental updates and several retrieval methods. Its self-managed deployment and stated requirements make it a fit for developers prepared to operate the framework and its supporting system.

Graphiti suits developers building agent systems that need changing facts, historical queries, provenance, or hybrid retrieval. It expects users to self-host and manage the surrounding system, with Python and an LLM API key available.

Overview

Graphiti is an open-source Python framework for building temporal context graphs that AI agents can query. It suits developers who need changing facts with traceable origins and are prepared to run the supporting system themselves. Its combination of historical queries, incremental updates and hybrid retrieval is compelling for that work, but it is not a managed service.

Key features

Graphiti tracks when facts become valid and when they are superseded, allowing applications to ask what is true now or what held at a past point in time. Incremental integration keeps new data flowing into the graph without rebuilding it from scratch, a useful fit for context that changes over time.

Entities and relationships derived from source episodes retain links to those episodes, which gives developers a route to inspect the context behind a result. Retrieval combines semantic search, keyword search and graph traversal, with result reranking. That range is useful when relevant information may be expressed in different ways or connected through relationships; teams still need to decide how those methods fit their application.

Developers can define entity and edge types with Pydantic models or allow structure to emerge from data. Graphiti supports Neo4j, FalkorDB and Amazon Neptune; Kuzu is deprecated. It defaults to OpenAI for LLM services and also supports Anthropic, Google Gemini, Groq and OpenAI-compatible providers through compatible endpoints. Python 3.10 or higher and an LLM API key are required, and structured-output-capable LLM services work best.

The project includes an MCP server for episode and entity management, relationship handling, search, group management and graph maintenance, along with a REST API service built with FastAPI. It can run locally, but users are responsible for deployment and operation. Anonymous usage statistics are collected by default and can be disabled with GRAPHITI_TELEMETRY_ENABLED=false; the project says telemetry excludes API keys, graph content, queries and personal identifiers. Vulnerability reports go through GitHub Private Vulnerability Reporting, and version 0.x is supported for security updates.

Pricing

Graphiti (open-source) costs nothing. The Apache-2.0 licensed framework is self-managed, so the price covers the software rather than the work of operating the surrounding system. There are no paid tiers or usage quotas in this plan; developers must bring an LLM API key and run the infrastructure themselves.

Platforms

Graphiti is available as an API and for self-hosted deployment. The Python requirement and LLM API key make it a framework for developers integrating it into their own systems, not a turnkey desktop or mobile app.

Who it's for

Choose Graphiti if you are building an AI agent that needs point-in-time context, incremental graph updates, provenance and several retrieval methods, and your team can manage the deployment. It is a poor fit if you need enterprise SLAs, support or security guarantees from the product maker: those are features of Zep's managed commercial platform.

Pros and cons

• Historical context: validity windows support queries about what was true at a given time.
• Traceable graph data: links to source episodes help developers follow the origin of derived entities and relationships.
• Flexible retrieval and integration: semantic search, keyword search, graph traversal and reranking pair with several supported graph databases and LLM providers.
• Operational responsibility: self-managed deployment means the team must build and run the surrounding system.
• Developer prerequisites: Python 3.10 or higher and an LLM API key are required, and structured-output-capable services work best.

Alternatives

Retrieval-Augmented Generation Tools is a broader category to explore when you want to compare options across RAG tools.

Choose AnythingLLM if you want a broader platform spanning Android, desktop, web and self-hosted use, with a free self-hosted Docker option and a paid Basic plan at $50 a month billed monthly. OpenRAG is another free self-hosted option, licensed AGPL-3.0 and described as having no feature held back; its managed service uses custom pricing.

PrivateGPT is a free open-source alternative for teams looking for API, desktop, web and self-hosted platforms. Quivr is worth considering for its free specialist agents, including documentation management and support-data analysis. Choose txtai for a free Apache-2.0 option installable with pip or Docker across API, desktop and self-hosted platforms.

Dify offers a more bounded free starting point: its Sandbox includes 200 message credits, one workspace, one member, five apps, 50 knowledge documents and 50MB of storage, among other caps. Haystack Open Source Framework is another free Python framework for API and self-hosted use. For a hosted workspace and pipeline tooling, Haystack Enterprise Platform has a free Studio plan with one workspace, one user, 100 pipeline hours and 50 files capped at 10MB each.

Verdict

Graphiti is a strong fit for developers building agent systems that need time-aware, traceable graph context and incremental updates. Its free Apache-2.0 framework and flexible retrieval are the main reasons to choose it; the need to operate the surrounding system is the main reason to look elsewhere.