Docs
Positioning
Source of truth for what DBX is. If a landing page contradicts this, the landing page is wrong.
DBX is the per-tenant memory engine for AI products: one isolated store per customer, holding both their working state and their vector memory.
Say / don’t say
| Say | Don’t say |
|---|---|
| Per-tenant memory engine for AI products | All-in-one AI database |
| Your existing RESP clients work | Drop-in Redis replacement |
| One isolated engine per customer | Multi-tenant namespaces |
| Cost per active tenant | Cheaper than Pinecone |
| Sized for per-tenant working sets | Scales to billions of vectors |
Who it is for
- AI agent platforms — private recall per end-customer
- Vertical AI SaaS — per-client RAG (legal, medical, support, finance)
- On-prem / regulated — one binary inside their network
- Teams running a cache and a vector DB and hand-rolling tenancy on both
Walk away if
- One tenant, one workload
- Billion-vector ANN, sharding, heavy filtering
- Fully managed vectors with no infrastructure
- SQL, joins, or a system of record
Five claims that have to stay true
- The tenant is a first-class object — directory, WAL, HNSW, snapshots.
- State and memory in one engine — one connection, one backup archive. Periodic
.rdbis KV-only; vectors live in.vecfiles. - Cost scales with active tenants — SQ8 mmap, idle tenants in page cache.
- Self-hosted single binary — embeddings never leave the network.
- Isolation Kernel — on Linux production a tenant is a sealed process with a Landlock-restricted filesystem, an encrypted WAL/checkpoint/index, and a Unix socket only the orchestrator can open. SQ8 rows stay mmap'd and unencrypted (use fscrypt); data in use is plaintext inside the worker.
Business
Self-host free, including inside your own SaaS. Commercial license to offer managed DBX to third parties. Converts to Apache 2.0 after four years.