AgentDB is a single-binary multimodal database — relational SQL, documents, vectors, full-text, graph, and git-style time-travel and branching in one engine, one transaction, one snapshot. Its flagship job is the flight recorder for AI systems: what your AI knew, what it retrieved, what prompt it used, what it decided, what policy applied — and any recorded run replays exactly, on demand, in SQL.

It includes one, but that is the smaller part. AgentDB is a single-node, crash-safe engine where SQL rows, vector search, native graph edges, branches, replay, evals, and a durable audit ledger all live in one snapshot. Vector search is a first-class column type queried inline with SQL — not a bolt-on index living beside your real database.

Postgres + pgvector gives you SQL and approximate vector search, then you assemble the rest yourself: a separate graph store, a vector cache, app-level versioning, and your own audit trail. AgentDB makes branches, AS OF time-travel, REPLAY, native graph traversal, eval gates, and a durable audit ledger native engine primitives over the same data — one snapshot, one transaction, one query surface.

Both are multi-model single-node engines, so the overlap is real. AgentDB is built for AI that must explain itself: branch a database, query AS OF any point in history, and REPLAY a past run deterministically — engine primitives, not app conventions. And at matched durability, on SurrealDB's own benchmark tool, AgentDB writes 5.2× faster than their flagship RocksDB backend — durability on, both sides.

Assist is the AI copilot docked inside AgentDB Studio. It grounds itself in your live schema, runs read-only queries instantly, and proposes any change as a confirmation card — nothing runs until a human approves it, and every action is audit-tagged. Nine providers work today, including Claude (sign in with your Claude account), OpenAI, Gemini, Bedrock, and Ollama fully local with no key at all. Provider keys live in your browser only, or sealed server-side in the vault — vault credentials are used by reference and never returned to the browser.

Yes — AgentDB Access is the built-in identity plane, no external provider needed. Humans sign in with passkeys, with MFA and recoverable single-use codes; agents get scoped keys. Every statement is authorized individually, refusals carry structured reasons, and credentials live in a sealed vault where agents use them by reference but can never read them. Break-glass recovery and emergency lockdown are built in.

AS OF reads the database at an exact point in history — a write number, a timestamp, or a named branch — so you can ask what a table looked like before a bad write. REPLAY OF takes a recorded run id and re-executes it deterministically against that historical state, so a flaky agent failure becomes a reproducible test you can step through instead of a screenshot you argue about.

Nothing acknowledged is ever lost — even when the machine dies mid-write. With the Lake tier configured, AgentDB keeps its whole history on object storage you own — your bucket, your keys. A machine can die mid-write; the next one attaches the same bucket and picks up exactly where it stopped, with fencing so exactly one writer holds the pen. You run it; there's no managed cloud yet.

Yes. Tenants get per-tenant isolation: separate logs, garbage collection, and checkpoints, with per-tenant limits so one caller can't starve the rest. Intra-tenant sharding is opt-in for hot tenants.

Production-ready for its shipped scope — self-hosted, a single-node writer over an object-storage shared commit log — GA since 2026-07-13 (v0.1.0; current release v0.1.2). Distributed it is not: one writer holds the pen. It is crash-safe — every commit is on disk before it's acknowledged, and committed data survives kill -9 even at relaxed durability. GA does not include: a managed cloud, multi-region replication, or a compliance certification. Encryption at rest is available (your KMS, envelope-wrapped keys), and the audit ledger is hash-chained and checkpoint-signed — compliance enablers you wire up, not a certified posture you inherit.

Several, all with conditions attached. Durable TPC-C: 10.69× SQLite-FULL and 4.43× Postgres on datacenter NVMe, durability matched at RPO=0 both sides (3.85× / 5.11× on laptop NVMe). Durable single-row writes: 2.6–2.8× RocksDB and 6.7–10× Fjall at matched RPO=0 (3-run cloud re-lock, error bars published). Branching: ~16× Dolt on AS OF point reads (embedded vs server). Vector search: 6.8× pgvector query speed at matched recall — and a statistical tie with Qdrant's stable runs (no Qdrant win claimed — its mean still edges ahead). Vector build: 2.4–3.0× pgvector; Qdrant keeps a 1.55–1.97× build lead — stated as a trade. What we don't claim: OLAP wins (DuckDB leads), a managed cloud, or a certification. Every claim: agentdb.run/claims — dated, SHA'd, reproducible.

The engine is proprietary (© Cingulum Technology). The SDKs and the published benchmark harness carry their own permissive licenses, so what runs inside your codebase stays unencumbered.