What is the best GraphRAG framework in 2026?

As of November 10, 2026: LightRAG (MIT, about 40,600 GitHub stars) for most document question-answering, because it is cheap to index, supports incremental updates and ships a web UI; Graphiti (Apache-2.0, v0.31.2) for agents whose facts change over time; the Neo4j GraphRAG Python package (Apache-2.0, v1.24.0) if you already run Neo4j; and Cognee (Apache-2.0, v1.6.7) for agent memory built on a knowledge graph. Microsoft GraphRAG is now in maintenance mode.

Is Microsoft GraphRAG still maintained?

Only for fixes. Microsoft's README, read November 10, 2026, says GraphRAG is largely in maintenance mode: it will not accept new pull requests or add features, and will ship bug fixes and dependency updates, especially for CVEs. The latest release is v3.3.0 from October 8, 2026. It remains a useful reference implementation, but new projects should prefer an actively developed framework.

When should I use GraphRAG instead of normal vector RAG?

Use GraphRAG when questions span many documents or depend on relationships: 'which suppliers are linked to these delayed projects', 'summarize the main themes across this corpus', or 'what did this customer prefer before and after March'. For simple lookups of a fact in one passage, vector search with a reranker is cheaper and usually as accurate. Graph extraction calls an LLM over every chunk, so indexing costs far more.

Which GraphRAG framework handles data that changes over time?

Graphiti. Each fact in a Graphiti context graph has a validity window recording when it became true and when it was superseded, every fact traces back to the raw episode it came from, and new data is added incrementally without recomputing the graph. LightRAG also supports incremental inserts and document deletion with automatic graph regeneration, but it does not model time.

Quick Answer

Best GraphRAG Frameworks in 2026: LightRAG, Graphiti, Neo4j

Published: November 10, 2026

The short answer

Start with LightRAG for document Q&A, Graphiti for agents with changing facts, and the Neo4j GraphRAG package if Neo4j is already your database. As of November 10, 2026:

LightRAG (HKUDS) — fast, cheap knowledge-graph RAG with a web UI and many storage backends. MIT.
Graphiti (Zep) — temporal context graphs for agents. Apache-2.0.
Neo4j GraphRAG for Python — Neo4j’s first-party, long-term-supported library. Apache-2.0.
Cognee — open-source agent memory on a self-hosted knowledge graph. Apache-2.0.
Microsoft GraphRAG — the original, now in maintenance mode. MIT.

The comparison

Versions, stars and licences from GitHub on November 10, 2026.

	LightRAG	Graphiti	Neo4j GraphRAG	Cognee	Microsoft GraphRAG
Latest release	v1.5.7 (Sep 2, 2026)	v0.31.2 (Oct 20, 2026)	1.24.0 (Nov 10, 2026)	v1.6.7 (Nov 6, 2026)	v3.3.0 (Oct 8, 2026)
GitHub stars	~40,600	~32,700	~1,300	~33,200	~36,600
Licence	MIT	Apache-2.0	Apache-2.0	Apache-2.0	MIT
Status	Active	Active	Active, vendor-supported	Active	Maintenance mode
Graph storage	Many: Neo4j, PostgreSQL, MongoDB, OpenSearch and more	Your graph database	Neo4j	Self-hosted, local by default	Files / Parquet
Updates	Incremental insert and delete	Incremental, time-aware	Pipeline per document	Incremental	Re-index
Best for	Document Q&A	Agents with changing facts	Teams on Neo4j	Agent memory	Research and global summaries

1. LightRAG — best default for document Q&A

LightRAG builds a knowledge graph from your documents and retrieves at two levels: specific entities and broader themes. It is cheaper to index than Microsoft’s approach, supports inserting and deleting documents with automatic graph regeneration, and in 2026 added a setup wizard, OpenSearch as a unified backend, four chunking strategies, separate LLMs per role (extraction, query, keywords, vision) and multimodal parsing via MinerU or Docling after merging RAG-Anything. It integrates RAGAS for evaluation and Langfuse for tracing.

Pick it if you want GraphRAG over PDFs, wikis or tickets with a UI your team can use on day one.

2. Graphiti — best for agents whose facts change

Graphiti, the graph engine inside Zep, builds temporal context graphs. Every fact carries a validity window (“as of March 2026”), entities have summaries that evolve, and everything traces back to the episode that produced it. Retrieval mixes semantic, keyword and graph traversal, and you can define your own entity and edge types with Pydantic. You supply the graph database, LLM and embedder; Zep’s hosted platform runs it for you if you prefer.

Pick it if your agent must know what is true now and what used to be true, such as customer preferences or account state.

3. Neo4j GraphRAG — best if Neo4j is already in your stack

Neo4j’s official Python package (Python 3.10–3.14) gives you a knowledge-graph builder pipeline and a set of retrievers: vector, hybrid, graph-traversal-enriched and Text2Cypher, which turns a question into a Cypher query. It works with OpenAI, Vertex AI, Ollama and other providers through optional extras, and Neo4j promises long-term support.

Pick it if your graph already lives in Neo4j or you want structured queries alongside semantic search.

4. Cognee — best for agent memory

Cognee turns documents, code and conversations into a self-hosted knowledge graph that agents search and reuse across sessions, and it can run locally on small CPU models with no API key. Custom data models and ontologies let you structure memory around your domain. See our agent memory systems ranking for how it compares with Mem0, Zep and Letta.

5. Microsoft GraphRAG — the reference, not the default

Microsoft’s GraphRAG introduced community summaries for “global” questions about a whole corpus, and it remains good for that. But its README now says the project is largely in maintenance mode with no new features, and it warns that indexing is expensive. Use it for research or if you already depend on it.

How to choose

Questions about documents? LightRAG.
Facts that change over time? Graphiti.
Already on Neo4j? Neo4j GraphRAG.
Long-term memory for agents? Cognee or Graphiti.
Not sure you need a graph? Try vector search plus a reranker first, and read RAG vs knowledge layer vs long context.

Graph extraction calls an LLM over every chunk, so index a small sample first and check the bill on our current API prices table.

Last verified: November 10, 2026.