One connected world.
Different ways to build it.

Embedded Python library. In-memory or persistent local storage; no database server.

Embedded analytical graph engine with Python and other language bindings.

Python applications use a network driver to connect to the DBMS.

Independent local processes can own read/write transactions on one store, with optimistic conflicts and commit-publication fencing.

Concurrent connections share one writable Database object. Direct file owners: one writable object OR multiple read-only objects. See the multiwrite note below.

Concurrent clients transact through the server. The DBMS coordinates locks and detects deadlocks.

ACID within the supported single-store contract. Snapshot reads and optimistic write-conflict validation; not universally serializable.

ACID transactions. The manual describes serializable execution; validate enabled multiwrite configuration separately.

ACID transactions with read-committed isolation by default. Explicit locks can strengthen isolation.

Ordered clauses, correlated OPTIONAL MATCH, EXISTS, bounded named paths and up to 64 UNION / UNION ALL branches. All 3,896 required V3 cases pass; one explicit upstream lists-of-maps divergence prevents a full conformance claim.

Cypher dialect with macros and typed graphs. The pinned 0.20.3 execution refused selected scoped CALL forms and produced different path multiplicities; neither dialect is established as a superset.

Labeled property graphs, Cypher, constraints and a broad procedure ecosystem. Dialects still differ.

Explicit typed tables coexist with flexible unlabeled nodes, dynamic properties and native ANY values. Node labels and relationship types may share a name; identities remain qualified by store, table and kind.

Structured property graph with declared node and relationship tables. Typed nested columns are supported; this is a different model from flexible, unlabeled entities.

Flexible labels and properties with constraints. It does not use Grafx-style typed table DDL or store arbitrary nested maps as properties.

Zero, one or multiple versioned labels on one stable node identity. CREATE/MERGE, SET/REMOVE, membership, retained history, copy and transfer/resume preserve labels. Physical typed constraints still apply.

The pinned 0.20.3 multiple-label scenario refused label SET. Existing node IDs stayed unchanged; this result concerns mutable labels, not all graph update support.

Native multiple labels and label updates through Cypher SET and REMOVE.

Returning and unit writing CALL, explicit/wildcard/leading-WITH imports, nested reads and updating UNION. Private branch effects share one outer statement rollback boundary; no inner commits.

The pinned run refused scoped CALL and the tested leading-WITH form. This does not prove absence of every CALL, UNION or subquery facility.

Returning and unit CALL subqueries with scoped variables and graph updates. Syntax and transactional forms remain version-dependent.

DATE, LOCALTIME, TIME, LOCALDATETIME, DATETIME and DURATION with constructors, clocks, arithmetic and durable native values. Equality keys are supported; ordered indexes for these families are refused.

Native DATE, timestamp families and INTERVAL. The pinned date scenario matched the expected calendar value; families and timezone semantics still require mapping.

Native date, local/zoned time and datetime, plus duration properties. Temporal values do not by themselves provide historical graph snapshots.

DECIMAL(p,s), up to 38 digits, with exact arithmetic, aggregates, ordering and typed equality-index seeks. Native history/copy and decimal128 columnar exchange preserve precision/scale. No decimal ordered index.

Native DECIMAL. The pinned storage scenario preserved 1.2500 as DECIMAL(12,4), using an explicitly declared CAST adaptation without a floating-point intermediate.

The documented native property types do not include a fixed-precision DECIMAL(p,s) type. INTEGER and FLOAT are not an exact decimal storage contract.

Typed LIST/MAP/ARRAY/STRUCT with nested nullability and exact descriptors, plus heterogeneous ANY values and lists of maps. Native, text and columnar consumers preserve supported types; collection property indexes are refused.

Nested LIST/ARRAY/STRUCT/MAP/UNION types. A pinned STRUCT[] persistence case matched Grafx LIST<STRUCT> with an explicitly equivalent declaration.

Homogeneous lists of simple property types can be stored. General maps, heterogeneous collections and lists containing null are not supported as stored properties.

Trusted typed read/write/unit procedures with native value/entity signatures, nested calls and explicit transaction-scoped query/schema authority. Revocation and shared budgets apply; external side effects are not rolled back.

Python and C++ user-defined functions are documented. An equivalent transaction-scoped writing/schema procedure authority contract was not established in this review.

Custom JVM procedures with READ, WRITE, SCHEMA and DBMS modes. Registration, permissions and execution contracts differ from embedded Python callbacks.

Native exact and HNSW search with explicit vector spaces, metrics, filters and resource bounds.

Vector extension with disk-based HNSW and filtered search.

Native ANN indexes for node/relationship embeddings. VECTOR property storage has separate format and edition conditions.

Native snapshot-bound weighted BM25 over node/relationship properties, with atomic analyzer/options replacement. New searches use the current analyzer with their owning data snapshot.

FTS extension with BM25 and configurable stemming and stop words.

Lucene-based full-text indexes with analyzers and optional eventually-consistent index updates.

Bounded prefix, exact phrases, persisted token positions and ordered proximity (slop). No phrase/prefix combination, character offsets, stemming or stopword configuration.

FTS extension with term matching, stemming and stop words. Equivalent typed slop and position-return contracts were not established by this review.

Lucene query syntax includes quoted exact matches and analyzers. Its visibility and returned evidence differ from the Grafx snapshot and token-position contract.

Native weighted RRF text/vector fusion with bounded graph evidence, on one snapshot and one target node table.

Vector and FTS extensions exist. An equivalent single-snapshot RRF plus graph-evidence API was not established in this review.

Query composition and the separate GraphRAG Python library provide HybridRetriever and HybridCypherRetriever.

Native bounded projections: BFS, Dijkstra, PageRank, WCC/SCC, k-core, label propagation and topological ordering with cycle detection.

Algo extension: PageRank, shortest paths, components, k-core and Louvain.

The separate GDS library provides extensive algorithms and ML pipelines, with its own edition limits.

Bounded CSV/JSONL/SQLite plus optional Arrow, Pandas, Polars, Parquet and NetworkX. Exact decimal and typed nested transport preserve descriptors with atomic imports; supported consumers and refusals are explicit.

CSV, Parquet, NumPy/DataFrames and extensions for external data systems.

Language drivers, import/export tools and integrations. GDS Arrow capabilities depend on its edition.

WAL, verified recovery, bounded physical backups and verified offline replacement restore. A physical replacement retains the database identity.

WAL, checkpoints and export. The reviewed material does not establish the same identity-aware restore contract.

CE supports offline dump/load. EE adds online backup and restore tooling.

Attached local catalogs, aliases, permissions and workspace resolution. Each transaction belongs to one store; no cross-store joins, edges or distributed COMMIT.

ATTACH / DETACH and external-system extensions. This review does not establish equivalent cross-store atomicity guarantees.

Additional databases and composite databases are Enterprise features, with a different ownership model from embedded local catalogs.

Resumable fresh-store transfer and bounded existing-target copy with durable receipts and optional endpoint closure. Qualified consumers preserve multiple labels, flexible/no-PK identities, relationship groups and expanded types. Historical transfer requires explicit current-only semantics.

CSV, Parquet and JSON export. Equivalent logical identity, durable receipts and resumption guarantees were not established by this review.

Dump/load and backup tooling provide different transfer contracts from the Grafx typed resumable logical-copy API.

Opt-in as-of reads, versions and graph/schema/property diffs; optional temporal index, retention, pins and quiescent physical compaction. Captures history from activation; no valid time or full bitemporality.

Temporal value types exist. Equivalent native retained graph versions and diffs were not established by this review.

CDC is available in Enterprise and specified Aura tiers. Change capture does not establish native arbitrary historical graph snapshots or equivalent diffs.

No native network service, replication or HA cluster. Direct local multi-process access is not replication.

Embedded ownership; a service can wrap the engine. Native HA clustering was not established by this review.

Client/server architecture. CE is single-instance; EE adds clustering and failover.

Host application and filesystem policy own access. No native users, RBAC, TLS service boundary or at-rest encryption.

Embedded database ownership is not an application user/RBAC system.

Server authentication; advanced RBAC and LDAP are Enterprise capabilities.

Python API, JSON CLI inventories/search, JS/TS subprocess recipe and script-free offline HTML snapshots. No native network driver or interactive GUI. Pulse supplies its own UI/MCP.

CLI, Ladybug Explorer and visualization integrations.

Neo4j Browser, Cypher Shell and additional tools; some are separate products.

Elastic License 2.0 plus the SaaS/Branding Addendum. Source available; review the actual terms.

MIT license.

Community: GPLv3. Enterprise: separate licensing.