The knowledge layer that makes your enterprise agent-ready.

Every AI system has a trust problem. Agents hallucinate. Data sources contradict each other. Nobody can trace where a "fact" came from or when it went stale. AttestDB is the truth layer underneath your AI — every fact stored with its source, confidence scored, contradictions caught, staleness flagged.

$ pip install attestdb

Free & open source. Full Rust engine. Runs locally forever.

How it works
Not a row. Not an edge. A claim.
Every database has an atomic unit. In a relational DB it's a row. In a graph DB it's an edge. In a vector DB it's an embedding. In Attest it's a claim — a fact with a source, a confidence score, and a timestamp. That one change makes everything else possible:
Relational row Graph edge Vector embedding Attest claim
Source tracking Optional Required on every write
Confidence Similarity score Evidence-weighted 0–1
Contradictions Last write wins Last write wins Last write wins Coexist until evidence resolves
Retraction Manual DELETE Manual One call, everything downstream flagged
Time travel Backup restore db.at(timestamp)
Self-correction
One bad source. Everything downstream, flagged.
The March load test had a bug in the test harness. In a normal system you'd spend a week asking around. In Attest you make one call.
db.retract_cascade("github:load-test-march-2024")
Retracted "Handles 10K req/s" github pr #847
Flagged "Use r5.4xlarge instances" confluence: capacity plan
Flagged "Selected Vendor B for CDN" jira: INFRA-2041
Safe "P99 latency < 50ms" production monitoring
retracted: 1 · flagged: 2 · safe: 1one call, zero archaeology
What changes
Static knowledge → Living knowledge
Most databases store what you put in and give back what you ask for. Attest goes further — it actively maintains the truth.

Freshness monitoring

Every claim decays over time. The system knows when your view of a customer is getting stale and flags it before you ask a bad question.

Pre-built answers

The background engine reads your query patterns and pre-synthesizes executive summaries for your most important entities. "Top 100 customers by risk" returns in milliseconds, not minutes.

Self-correction

Retract one bad source and everything downstream is automatically flagged. Facts corroborated by independent sources survive. One call, zero archaeology.

Contradiction tracking

When two sources disagree, both claims coexist with confidence scores. The system doesn't silently pick one — it shows you the tension and the evidence.

Use cases
One knowledge layer. Many applications.
For coding agents

Persistent Brain

Give Claude Code, Cursor, or any MCP agent a persistent brain that tracks bugs, patterns, and dead ends across sessions.

Learn more →
For teams

Commitment Tracker

Track commitments across Slack, email, and meetings. Detect stalling projects and draft supportive check-ins.

Learn more →
For enterprise

Data Unification

Connect 5 Salesforce instances and 2 SAP systems. AI maps schemas, resolves entities, enforces access control. Cross-source questions answered in seconds.

Learn more →
Under the hood
Ingestion
1.3M claims/sec
Rust engine, single-threaded
Query latency
~12µs indexed lookups
Microsecond queries, sub-millisecond on LMDB
Storage
Single file
No server, no config — like SQLite
Integrations
30 connectors + 106 MCP tools
Slack, GitHub, Gmail, Postgres + Claude Code, Cursor, Windsurf, Codex, Gemini
Why Attest
Not another database. A different primitive.
Vector DBs find similar text. Graph DBs store edges. AttestDB tracks whether claims are true.
AttestDB Pinecone / Weaviate Neo4j PostgreSQL
Atomic unit Sourced claim Vector embedding Edge Row
Provenance Required on every write Optional metadata Optional property Not built-in
Retraction cascade Automatic Manual Manual
Contradiction handling Evidence-weighted Last write wins Last write wins Last write wins
Confidence scoring Built-in (0–1) Similarity score
Query latency ~12µs ~10ms ~5ms ~1ms
MCP tools 106
We were running dozens of agents across thousands of research papers. They extracted findings, made claims, contradicted each other — and we had no way to store what they collectively knew in a form anyone could reason about.

Built at Omic, a computational biology company where AI agents ingest thousands of papers daily. The vector databases lost provenance. The graph databases couldn't handle contradictions. The document stores couldn't be queried structurally. So we built the layer underneath. The problem isn't unique to biology — it's everywhere knowledge is fragmented across systems and held together by human memory.

Start in 60 seconds. No account required.

$ pip install attestdb && attestdb quickstart

Free & open source. Full Rust engine. Runs locally forever.

Try the live demo Read the docs