Slater packs massive graph DBs into a few hundred MB

A new open-source project called Slater is aiming at one of graph databases' most persistent problems: memory use. Its pitch is simple: serve graphs that do not fit in memory — including hundreds of millions of nodes and billions of edges — while keeping RAM usage in the low hundreds of MB, not tied to graph size.

Slater exposes graphs over standard Bolt, so existing Neo4j drivers work without modification. According to the project, the database can serve the 90 million node / 1.5B-edge Wikidata graph from a few hundred MB of RAM by paging from an on-disk image instead of keeping the full graph resident. By contrast, the README says in-memory graph engines may need ~64–128 GiB to open that dataset.

The architecture splits graph building from graph serving. An offline tool, slater-build, compiles data into an immutable, content-addressed on-disk image. The online slater server then serves that image with a fixed cache budget, meaning a 4 GB graph and a 400 GB graph can use the same amount of RAM at runtime.

Slater is not read-only, though writes are optional. The project uses an LSM-based writable layer on top of the immutable core, with a write-ahead log, an in-memory table, immutable delta segments, and periodic consolidation back into a fresh core. The stated goal is to keep the read path unchanged when there are no pending writes, while still allowing durable updates over Bolt.

Recommended reading

Microsoft taps Mistral in multi-billion AI infrastructure deal

Key features highlighted by the project include:

• Bolt 5.4 / 4.4 / 4.1 compatibility
• A broad Cypher query surface plus a subset of ISO GQL (ISO/IEC 39075)
• Built-in disk-native vector search using Vamana + PQ with cosine, L2, or dot KNN
• Storage on local filesystems, S3, or Google Cloud Storage
• Optional XChaCha20-Poly1305 encryption at rest and TLS in transit
• Multi-tenant support and ACLs

The current version listed in the repository is v0.24.1. For teams building knowledge graphs, recommendation systems, identity graphs, or graph-backed retrieval for RAG, Slater’s core claim is unusually direct: graph size should live on disk, not in your memory bill.