Graph database. Or memory platform.
Neo4j is the category-defining graph database, now with GenAI features layered on. HangarX Cortex is a memory platform built on a graph database — extraction, claims, hybrid retrieval, contradiction detection, and an MCP server, bundled.
Different layers of the same stack. Many teams use Neo4j as their graph DB and add memory infrastructure on top — which is what Cortex is.
Pick Cortex if
You want managed memory infrastructure, not a graph database.
Extraction, claims, retrieval, reranking, contradiction detection, MCP serving — bundled. You ingest a corpus and your agents start querying.
Pick Neo4j if
You need a billion-scale graph database with mature enterprise tooling.
Graph Data Science library, decades of operational maturity, native vector indexes, and the deepest catalog of graph algorithms in the industry.
What we agree on
- Graphs are the right shape for AI memory — entities and relationships unlock multi-hop reasoning
- Cypher is the right query language for property graphs
- Vector + graph beats either alone — both support hybrid retrieval
- Self-host should be a first-class option for regulated workloads
- GraphRAG patterns are the future of grounded retrieval
- Enterprise-grade governance and certifications matter for some buyers
Where we differ
| Dimension | HangarX Cortex | Neo4j |
|---|---|---|
| Layer of the stack | Memory platform | Graph database (with GenAI features layered on) |
| Cypher query language Cortex uses FalkorDB which is Cypher-compatible. Neo4j is the original Cypher implementation with deeper feature coverage. | ||
| Graph DB scale ceiling | Tens of millions of nodes | Billions of nodes |
| Native vector index inside the graph | ||
| LLM extraction (entities, relationships, claims) Neo4j's GraphRAG / GenAI features include extraction primitives, but you assemble the pipeline. Cortex ships extraction with claim provenance as a built-in step. | ||
| Document parsing + chunking | ||
| Hybrid retrieval pipeline (BM25 + vector + graph + reranker) | ||
| Claims with provenance (SPO triples + source spans) Neo4j stores graph data. Provenance is something you'd model and query yourself. | ||
| Contradiction detection across claims | ||
| MCP server (cross-tool memory) | ||
| Native Obsidian plugin | ||
| Self-host | ||
| Enterprise / regulated industries posture Neo4j has decades of enterprise procurement, certifications, and governance. Cortex is newer; the local Docker stack lets you run inside already-compliant environments. | ||
| Graph algorithms (PageRank, community detection, etc.) Cortex uses Personalized PageRank for retrieval expansion. Neo4j's Graph Data Science library has a much broader algorithm catalog. |
The wedge
Neo4j is the gold standard for graph databases. Decades of operational maturity, the deepest graph algorithm catalog (Graph Data Science library), and a real GenAI roadmap. If you have an existing Neo4j deployment and graph engineers, building agent memory on Neo4j is a defensible choice.
But a graph database is not the same as a memory platform. Neo4j gives you the storage and the query language. To turn that into agent memory, you still build: parsing, chunking, LLM extraction with claim provenance, hybrid retrieval (BM25 + vector + graph), reranking, contradiction detection, and an MCP serving layer. That's the shape of Cortex — the memory platform built on a graph database, instead of asking you to build it on one.
When you should pick Neo4j
- You're operating at billion-node graph scale with strict performance SLAs.
- You need the broadest catalog of graph algorithms (PageRank, community detection, centrality, etc.) via the Graph Data Science library.
- Your enterprise procurement requires a vendor with decades of compliance certifications and global operational maturity.
- Your team has graph engineers and an existing Neo4j deployment you want to extend with GenAI features.
- Your workload is general-purpose graph analysis, not specifically agent memory.
When you should pick Cortex
- You want grounded agent memory shipping in days — extraction, retrieval, and serving already wired up.
- Your problem is question-answering over a corpus, not general-purpose graph analysis.
- You'd rather not build parsing, extraction, retrieval, reranking, contradiction detection, and MCP serving on top of a graph DB.
- You want MCP-native serving so Claude, Cursor, Cline, and Windsurf all share one memory.
- You're an Obsidian user and want a native plugin.
FAQ
Can I use Neo4j as the graph backend for Cortex?
Cortex defaults to FalkorDB (Cypher-compatible, Redis-protocol). FalkorDB is faster for many memory workloads and ships in a single Docker container. If your team is already standardized on Neo4j and you want graph queries to land there, Cortex's storage layer is open enough that pointing it at Neo4j is feasible but not the default path. Most teams pick whichever graph DB their ops team prefers.
Why not just build agent memory directly on Neo4j?
Many enterprise teams do — and many spend the next year building parsing, extraction, claim modeling, retrieval pipelines, reranking, contradiction detection, and serving infrastructure on top. That work is roughly what Cortex has already shipped. Neo4j gives you a phenomenal graph database; Cortex gives you the memory platform built on a graph database. Different layers of the same stack.
What about Neo4j's GraphRAG / GenAI features?
Neo4j has been adding meaningful GenAI capabilities — vector indexes inside the graph, LLM extraction primitives, GraphRAG patterns. They're moving up the stack. The honest difference today: Cortex's pipeline is opinionated and end-to-end (parsing → extraction → claims → hybrid retrieval → reranking → MCP serving), while Neo4j's GenAI features are primitives you compose. If your team has graph-DB engineers and a long roadmap, Neo4j gives you ceiling. If you want grounded memory shipping fast, Cortex is closer to ready.
Which scales further on graph queries?
Neo4j wins on raw graph scale — billions of nodes with mature graph algorithms (Graph Data Science library), enterprise-grade replication, and decades of operational tooling. Cortex (FalkorDB) handles tens of millions of nodes well but isn't engineered for the billion-node tier. If your knowledge graph alone is gigantic and graph algorithms are the workload, Neo4j is the right call.
Related comparisons
If you're evaluating this against Cortex, you're probably also weighing these.
A graph DB is the foundation. Cortex is the platform.
Skip the build-it-yourself memory pipeline. Cortex bundles extraction, retrieval, claims, and MCP serving on top of a Cypher-compatible graph store.