Graph Database
for AI Agents,
Knowledge Graphs,
and GraphRAG

Q: What is FalkorDB?

FalkorDB is an open source graph database built for AI. It stores graphs as sparse matrices and runs traversals as linear algebra, which keeps multi-hop queries fast. Teams use it for knowledge graphs, GraphRAG, AI agent memory, security graphs, and fraud detection.

Q: Is FalkorDB open source?

Yes. The full engine, including multi-graph and multi-tenancy, is on GitHub. FalkorDB Cloud and Enterprise add managed operations, high availability, and support on top of the same engine.

Q: What query language does FalkorDB use?

FalkorDB uses openCypher, the most widely adopted graph query language. Native Cypher support means migrating to FalkorDB takes hours, not days, and migration guides are available in the docs.

Q: How does FalkorDB compare to other graph databases?

FalkorDB is built for low-latency traversal in AI applications. It stores graphs as sparse matrices, runs thousands of isolated graphs per deployment, uses openCypher, and keeps the full engine open source. For a side-by-side comparison, read FalkorDB vs Neo4j for AI applications. This page shows how graph, relational, and vector databases differ.

Q: Can I migrate from another graph database to FalkorDB?

Yes. Queries written in Cypher carry over with little change, so migrating to FalkorDB takes hours, not days. The docs include step-by-step migration guides, and an engineer can review your schema and queries before you move.

Q: What is the best graph database for AI agents and GraphRAG?

Look for low latency at depth, vector search in the same query as traversal, isolated graphs per user or tenant, and framework support. FalkorDB covers all four, and its GraphRAG-SDK ranks #1 on the GraphRAG-Bench Novel and Medical leaderboards.

Q: Does FalkorDB support vector search?

Yes. FalkorDB has native vector indexes, so one query can find semantically similar content and then traverse the relationships around it. This is the core pattern behind GraphRAG and hybrid retrieval.

Q: Graph database or vector database for RAG?

Use vectors to find similar text and a graph to follow how facts connect. Questions that span several entities or documents need relationships, which vector search alone cannot return. FalkorDB runs both in one database.

Q: What is the fastest graph database?

Speed depends on the workload, so test with your own data. FalkorDB is built for low-latency traversal: it delivers 496× faster P99 latency than a leading graph database in our open benchmark. The dataset, queries, and code are public at benchmark.falkordb.com.

Q: Is FalkorDB an in-memory database? Is my data durable?

FalkorDB serves the graph from memory for low latency and persists it to disk with snapshots and append-only files. Paid Cloud plans add automated backups, and Pro and Enterprise add continuous persistence and high availability.

An enterprise-grade graph database that answers multi-hop queries in milliseconds and runs thousands of isolated graphs in a single deployment.

Talk to an engineer Start free

Trusted by

How FalkorDB works

FalkorDB stores your graph as sparse matrices and answers multi-hop queries with matrix multiplication, so results come back in milliseconds.

Step 01

Data ingestion

Load data from tables, documents, and events.

Step 02

Stored as a graph

Data is stored as nodes and edges. Under the hood, FalkorDB represents the graph as sparse matrices, which keeps memory low and lookups fast.

Step 03

Queried in milliseconds

You query in openCypher. Traversals run as linear algebra on those matrices, so multi-hop answers come back with very low latency.

Sparse matrix engine

FalkorDB represents graph topology as sparse adjacency matrices and executes traversals as linear algebra operations. Multi-hop queries return with consistent low latency as data volume and query depth grow.

In-memory execution with durable storage

The full graph is served from memory. Snapshots and append-only files keep data durable on disk.

Multi-tenant graphs

Each customer, agent, or session's data lives in a dedicated, isolated graph. Every query is scoped to exactly one graph, so tenant isolation is enforced at the architectural level.

Vector, full-text, range, and array indexes

Combine graph traversal with semantic search in the same query, turning FalkorDB into the ultimate backbone for your AI applications.

Graph algorithms in the database

Run shortest paths, betweenness centrality, connected components, community detection, max flow, harmonic centrality, and more, directly in the database.

User-defined functions

Extend and customize Cypher with your own logic, plus the FLEX library of ready-made functions.

Try it on your machine or in the cloud

Running in 30 seconds. No credit card required for Cloud.

Start free Read the docs →

terminal

$ docker run -p 6379:6379 -p 3000:3000 -it --rm falkordb/falkordb:latest

Why teams choose FalkorDB

Built for AI

Combine graph traversal with semantic search in a single query to build knowledge graphs, GraphRAG pipelines, and agentic applications on one platform.

Performance

An engine built for enterprise-grade workloads that doesn't slow down as graphs grow.

Scalability

Scale out by distributing graphs across shards and adding read replicas, so throughput grows with your cluster.

Multi-tenant graphs

Host thousands of tenants in a single deployment. Each customer, agent, or session's data resides in its own graph, and every query is scoped to one graph, so isolation is enforced by the architecture with no impact on performance.

Enterprise ready

Production-ready at enterprise scale: horizontal scaling, built-in multi-tenancy, high availability with multi-zone deployment, and SOC 2 Type II certification. Deploy fully managed, in your own cloud account, or on-premises, backed by dedicated engineering support and architecture consulting.

Engineering foundations

Rust powered Stable by design, no surprises.

Open source Free to use, free to fork.

OpenCypher No new query language to learn. Migrating takes hours, not days.

See what our graph database can do.

An open benchmark against Neo4j. The dataset, the queries and the code are public, so you can run it yourself.

Explore the live benchmark

P99 latency

496

faster

FalkorDB

83 ms

Neo4j

41,157 ms

Memory

6.3

leaner

FalkorDB

95 MB

Neo4j

600 MB

Throughput

11.4

FalkorDB

2,500 queries/sec

Neo4j

220 queries/sec

Lower latency, lower memory and higher throughput are better.

P99 latency

496 ×

faster

FalkorDB 83 ms

Neo4j 41,157 ms

Memory

6.3 ×

leaner

FalkorDB 95 MB

Neo4j 600 MB

Throughput

11.4 ×

FalkorDB 2,500 queries/sec

Neo4j 220 queries/sec

Lower latency, lower memory and higher throughput are better.

Open source

free to use, free to fork

Redis based

runs on infrastructure you already trust

OpenCypher

no new query language to learn

Rust powered

stable by design, no surprises

What can you build with a graph database?

AI applications

I’m building with LLMs and agents.

AI applications

I’m building with LLMs and agents.

Graph workloads

I’m querying connected data.

Graph workloads

I’m querying connected data.

For AI teams that need answers grounded in how things connect, and memory that lasts past one session.

GraphRAG

Ground LLM answers in connected context, not just similar text.

Question

"Which open claims are covered by the policy we changed last week?" # graph query generated MATCH (p:Policy)-[:COVERS]->(c:Claim)-[:FILED_BY]->(u:Customer) WHERE c.status = 'open' AND p.updated_at >= date() - duration('P7D') RETURN c.id, u.name

Answer grounded in 3 connected facts, each with its source

GraphRAG

Ground LLM answers in connected context, not just similar text.

Agentic AI flows

Give agents a shared, queryable model of the world they act in.

Agent tool call 3 ms

MATCH (t:Ticket {id: 4821})-[:ABOUT]->(s:Service) -[:DEPENDS_ON]->(d:Service) RETURN d

2 upstream services returned to the agent as its next step

Agentic AI flows

Give agents a shared, queryable model of the world they act in.

Agentic memory

Long-term memory that remembers relationships, not only facts.

Cypher 2 ms

MATCH (u:User {id: 'u_482'})-[:IN]->(s:Session) -[:MENTIONED]->(t:Topic) WHERE s.id <> $session_id RETURN DISTINCT t

2 topics recalled from 2 earlier sessions

Agentic memory

Long-term memory that remembers relationships, not only facts.

Question

Answer grounded in 3 connected facts, each with its source

Agent tool call 3 ms

MATCH (t:Ticket {id: 4821})-[:ABOUT]->(s:Service) -[:DEPENDS_ON]->(d:Service) RETURN d

2 upstream services returned to the agent as its next step

Cypher 2 ms

MATCH (u:User {id: 'u_482'})-[:IN]->(s:Session) -[:MENTIONED]->(t:Topic) WHERE s.id <> $session_id RETURN DISTINCT t

2 topics recalled from 2 earlier sessions

AI frameworks integrate out of the box, from LangChain and LlamaIndex to Graphiti and Mem0

GraphRAG-SDK on GitHub → Read the docs →

7 AI frameworks integrate out of the box, from LangChain and LlamaIndex to Graphiti and Mem0

GraphRAG-SDK on GitHub → Read the docs →

For fraud, security and supply-chain teams that need answers across many hops, in real time.

Fraud & risk detection

Follow the money across accounts, devices and identities in real time.

Cypher 3 ms

MATCH (a:Account {flagged: true})-[:USES]->(:Device) <-[:USES]-(b:Account) WHERE a <> b AND NOT b.flagged RETURN DISTINCT b

2 accounts share a device with a flagged account

Fraud & risk detection

Follow the money across accounts, devices and identities in real time.

Cybersecurity

Map who can reach what: identities, permissions, assets and the paths between them.

Cypher 4 ms

MATCH p = (u:User {role: 'contractor'}) -[:MEMBER_OF|CAN_ACCESS*1..4]->(a:Asset {tier: 'critical'}) RETURN p

1 path from a contractor account to a critical asset

Cybersecurity

Map who can reach what: identities, permissions, assets and the paths between them.

Supply chain

Trace dependencies upstream and downstream before a disruption becomes an outage.

Cypher 4 ms

MATCH (p:Port {closed: true})<-[:SHIPS_VIA]-(:Carrier) <-[:USES]-(s:Supplier) RETURN DISTINCT s

3 suppliers ship through the closed port

Supply chain

Trace dependencies upstream and downstream before a disruption becomes an outage.

Cypher 3 ms

MATCH (a:Account {flagged: true})-[:USES]->(:Device) <-[:USES]-(b:Account) WHERE a <> b AND NOT b.flagged RETURN DISTINCT b

2 accounts share a device with a flagged account

Cypher 4 ms

MATCH p = (u:User {role: 'contractor'}) -[:MEMBER_OF|CAN_ACCESS*1..4]->(a:Asset {tier: 'critical'}) RETURN p

1 path from a contractor account to a critical asset

Cypher 4 ms

MATCH (p:Port {closed: true})<-[:SHIPS_VIA]-(:Carrier) <-[:USES]-(s:Supplier) RETURN DISTINCT s

3 suppliers ship through the closed port

496×

faster P99 latency than Neo4j, on an open benchmark

See the live benchmark → Read the docs →

496× faster P99 latency than Neo4j, on an open benchmark

See the live benchmark → Read the docs →

What is a graph database?

A graph database stores data as nodes (entities), edges (relationships), and properties (attributes). Relationships are saved as part of the data, not rebuilt at query time with joins. Questions that follow connections, such as fraud rings, supply chains, or how two concepts relate, stay fast and simple to write.

FalkorDB

FalkorDB is an open source graph database built for AI. It represents graphs as sparse adjacency matrices and runs traversals as matrix multiplication using GraphBLAS. Queries are written in openCypher. FalkorDB runs fully managed on FalkorDB Cloud, inside your own AWS or Google Cloud account, or self-hosted.

node

edge

property

Graph database vs relational database vs vector database

Compare with

Graph database

Relational database

Vector database

Data model Nodes, edges, properties Tables, rows, columns Embeddings (numeric vectors) with metadata

How it finds answers Follows relationships between entities Joins tables on matching keys Returns the items closest in meaning to the query

Relationships Stored with the data Rebuilt with joins at query time Not stored; similarity is computed per query

Multi-hop queries Stay fast as depth grows Slow down with each added join Not supported

Schema changes Add labels and properties without migrations Migrations and query rewrites Re-embed content when the model or chunking changes

Best for Connected data, traversals, AI retrieval Reporting, ledgers, aggregations Semantic search over unstructured text

Read the beginner’s guide: Graph Database Explained → Compare FalkorDB with other graph databases →

Run it locally. Ship it anywhere.

Running in 30 seconds. No new language to learn.

Quickstart and installation

→

Documentation

→

Client libraries Python · Node · Java · Go · Rust

→

Start building

Additional tools

FalkorDB Browser Visualize, query, and inspect graphs in a web UI.

Open the Browser →

FalkorDBLite An embedded, zero-config FalkorDB for Python and TypeScript, for tests, prototypes, and ephemeral graphs.

Read the docs →

Run it where your data lives

Managed by FalkorDB

FalkorDB's cloud

FalkorDB Cloud

Fully managed on AWS, Google Cloud, and Azure, with multi-tenancy on every plan.

Start free →

Your cloud

Bring your own cloud

Fully managed FalkorDB inside your own cloud account. Your data never leaves your subscription.

Talk to an engineer →

Run by you

Your servers

Self-hosted

Run the open source engine with one Docker command, or self-host Enterprise with a management console.

View on GitHub →

Your application

Embedded and warehouse

FalkorDBLite runs in-process for Python and TypeScript. FalkorDB on Snowflake queries warehouse tables as graphs.

Read the docs →

← Less to operate Closer to your data →

Graph database FAQ

The basics

What is FalkorDB?

FalkorDB is an open source graph database built for AI. It stores graphs as sparse matrices and runs traversals as linear algebra, which keeps multi-hop queries fast. Teams use it for knowledge graphs, GraphRAG, AI agent memory, security graphs, and fraud detection.

Is FalkorDB open source?

Yes. The full engine, including multi-graph and multi-tenancy, is on GitHub. FalkorDB Cloud and Enterprise add managed operations, high availability, and support on top of the same engine.

What query language does FalkorDB use?

FalkorDB uses openCypher, the most widely adopted graph query language. Native Cypher support means migrating to FalkorDB takes hours, not days, and migration guides are available in the docs.

How does FalkorDB compare to other graph databases?

FalkorDB is built for low-latency traversal in AI applications. It stores graphs as sparse matrices, runs thousands of isolated graphs per deployment, uses openCypher, and keeps the full engine open source. For a side-by-side comparison, read FalkorDB vs Neo4j for AI applications. This page shows how graph, relational, and vector databases differ.

Can I migrate from another graph database to FalkorDB?

Yes. Queries written in Cypher carry over with little change, so migrating to FalkorDB takes hours, not days. The docs include step-by-step migration guides, and an engineer can review your schema and queries before you move.

AI and RAG

What is the best graph database for AI agents and GraphRAG?

Look for low latency at depth, vector search in the same query as traversal, isolated graphs per user or tenant, and framework support. FalkorDB covers all four, and its GraphRAG-SDK ranks #1 on the GraphRAG-Bench Novel and Medical leaderboards.

Does FalkorDB support vector search?

Yes. FalkorDB has native vector indexes, so one query can find semantically similar content and then traverse the relationships around it. This is the core pattern behind GraphRAG and hybrid retrieval.

Graph database or vector database for RAG?

Use vectors to find similar text and a graph to follow how facts connect. Questions that span several entities or documents need relationships, which vector search alone cannot return. FalkorDB runs both in one database.

Performance and scale

What is the fastest graph database?

Speed depends on the workload, so test with your own data. FalkorDB is built for low-latency traversal: it delivers 496× faster P99 latency than a leading graph database in our open benchmark. The dataset, queries, and code are public at benchmark.falkordb.com.

Is FalkorDB an in-memory database? Is my data durable?

FalkorDB serves the graph from memory for low latency and persists it to disk with snapshots and append-only files. Paid Cloud plans add automated backups, and Pro and Enterprise add continuous persistence and high availability.

How does FalkorDB scale?

FalkorDB scales out by distributing graphs across shards and adding read replicas. In a multi-tenant permissions workload, throughput grew from about 25,000 to 120,000 queries per second when moving from one node to three primaries with three replicas.

How does FalkorDB handle multi-tenancy?

Each customer, agent, or session runs in its own isolated graph, with no impact on performance. A single deployment can hold thousands of graphs, and multi-tenancy is included on every plan, including Free.

Getting started, pricing and security

How do I get started with FalkorDB?

Run docker run -p 6379:6379 -p 3000:3000 -it --rm falkordb/falkordb:latest and open the browser at localhost:3000, or sign up for the free tier on FalkorDB Cloud. Either way you are running in about 30 seconds, with no new language to learn.

How much does FalkorDB cost?

You can start free on FalkorDB Cloud or self-host the open source engine. Paid Cloud plans start at $73 per month for 1 GB, Pro starts at $350 per month for 8 GB, and Enterprise pricing is tailored.

Is FalkorDB secure enough for enterprise use?

Yes. FalkorDB is SOC 2 Type II certified and supports TLS, VPC isolation, and graph-level access control. Enterprise customers can also run it inside their own cloud account.

Can I run FalkorDB locally?

Yes. Run FalkorDB with a single Docker command, or use FalkorDBLite, an embedded version for Python with no server to manage. When you are ready, move the same graphs to FalkorDB Cloud or your own infrastructure.