One engine. Every modality. Measured.
AgentDB is four products on one engine. The Engine holds every shape of data — rows, documents, vectors, search, graph — in one transaction, with one history. The Lake keeps that history on object storage you own. Access is the built-in front door: passkeys for people, scoped keys for software, every decision signed. Studio is the console you can talk to — ask in plain English, watch it work, on the record. And because the history is the data, every answer comes with why.
The problem
AI stopped suggesting and started deciding — loans, diagnoses, refunds, code. Its memory today is scattered across nine systems: a database, a vector store, a graph DB, a prompt registry, an eval service, a trace store, an audit log, an access policy, a workflow engine. When someone asks why did it do that, the answer is reassembled from all nine — or it isn’t. When the answer matters, “we can’t reconstruct it” isn’t an answer.
Proven — Wave 5 + Clean Sweep · July 2026
| Claim | Figure | Condition | Verified |
|---|---|---|---|
| Durable OLTP throughput (TPC-C) | 10.69× | durability matched — RPO=0 both sides, fsync-before-ack · datacenter NVMe (Hetzner ccx33) · 22 warehouses, 22 terminals | 2026-07-04 · c52131a |
| Durable writes (crud-bench, 9 competitors) | 5.2× | durability on for every arm (--sync) · 100k rows, 12 clients, 24 threads · ccx33 · AgentDB 97,876 vs 18,705 creates/s. Also: 6.0× Postgres, 29.8× MySQL, ~97× SQLite-FULL | 2026-07-04 · c52131a |
| AS OF point reads (branch time-travel) | ~16× | 10k rows, 1,000 branches, RPO=0 both arms, same box, arms interleaved · AgentDB 0.003 ms vs Dolt 0.052 ms p50 | 2026-07-07 · 3a02f7e |
| Vector ANN query (SIFT-1M, real corpus) | 4.5× | 1M vectors, 10k queries, HNSW params matched (M=48, ef=500) · recall@10 0.9986 vs 0.9992 · p50 2.40 ms vs 10.86 ms | 2026-07-07 · 220f09a |
| Multi-modal atomicity | one transaction | proven kill-9 durable (Wave 1, 2026-07) — the only engine demonstrating this | 2026-07-01 · e56a61f |
The losses are published too — OLAP (DuckDB leads), vector index build, a query gap to Qdrant. Every row, including those, lives in the claims ledger with a runnable command. Benchmarks run in numbered waves; the 2026-07-12 cloud re-lock put 3-run error bars on the standing ratios. The next wave re-runs the full set — verdicts may change; the dates and SHAs never do.
What it is
One engine — rows, vectors, full text, and graph committed in one transaction, one snapshot; every commit on disk before it’s acknowledged. One language — AgentSQL: it’s SQL, with six verbs SQL never had (branch, time-travel, replay, eval-gate, deploy, blend). One console — Studio, the control room, with Assist: ask in plain English, it shows its work and asks before it changes anything. Built-in identity — AgentDB Access: passkeys for humans, scoped keys for agents, a sealed credential vault. Native MCP — the socket your agents already speak, through the same gates as everything else. Every door runs the same conformance suite.
| Name | What it means |
|---|---|
| AgentDB | The product — the umbrella brand (the platform name) |
| AgentSQL | The query language — SQL, with six verbs SQL never had |
| AgentDB Studio | The console — the control room, with Assist docked |
| AgentDB Access | The built-in identity & governance plane |
| AgentDB Engine | Storage + compute — one Rust binary |
| AgentDB Lake | The object-storage tier — your bucket, your keys, the whole history |
agentsql-* | The binary and crate prefix (agentsql-server, agentsql CLI) — named for the query language |
Where it is today
GA since 2026-07-13 (v0.1.0; current release v0.1.2) for the self-hosted, single-node-writer scope. The crash suites run with real SIGKILL in the executable release gate on every merge. No managed cloud — we deploy it with you, on your machines and your object storage. Closed source — the benchmark harness is public, and we’ll run any number on this page in front of you, on hardware you pick. One founder — you work directly with the person who built the engine.
Who it’s for
Engineers replacing the nine-system stack with one engine. Teams whose AI answers to a regulator — healthcare, finance, law. Anyone who’s been told “the computer decided” and wanted to see the receipts.
The ask
We’re taking design partners: we deploy AgentDB into your stack with you, and your hardest “prove it” requirement shapes the engine. Apply here — five fields, two minutes. We reply within two working days.
AgentDB exists because too many systems answer “we can’t reconstruct it” — and I wanted to build one that never has to. So the engine can’t give that answer: every commit is on disk before it’s acknowledged, every decision replays exactly, and even “no” goes in the ledger. The discipline is the product. The crash suite runs with real SIGKILL in the blocking release gate on every merge; competitors are raced with durability settings matched; the losses are published next to the wins. If a number on this site stops being true, it changes here first. One founder means you work directly with the person who built the engine — ask me anything, and I’ll show you the receipts.
— Sam Ginsberg
Ship notes
A short note when something real ships — wave results, new surfaces, changed verdicts. Archived publicly at /notes.
Numbers: Wave 5 + Clean Sweep · July 2026, from the public benchmark ledger — each with date, engine SHA, and reproduction command at /claims. Harness: github.com/meshbergio/agentdb-bench-harness.