AI agents are getting powerful. They can write code, analyze data, manage projects, and make decisions. But they all share the same weakness: they can't tell what's true.
Ask an agent a question and it gives you an answer. But where did that answer come from? Is it still current? What if two sources disagree? What if the underlying data was retracted last week? The agent doesn't know, and neither do you.
This isn't a retrieval problem. It's a truth problem.
The industry has tried to solve it with bigger context windows, better embeddings, and fancier RAG pipelines. But these all address the same question: "what information is relevant?" None of them address the harder question: "is this information still true?"
Truth requires provenance. Every fact needs a source. Not as optional metadata — as a structural requirement. A fact without a source is an opinion.
Truth requires confidence. Three independent sources confirming the same fact is stronger than one source at high confidence. A fact confirmed yesterday is stronger than one confirmed last year.
Truth requires contradiction tolerance. Reality is messy. Two sources will disagree. A system that silently picks one answer is lying about its certainty. The honest response is to show both claims and the evidence behind each.
Truth requires self-correction. When a source is retracted — a bad dataset, a buggy sensor, a dishonest report — everything downstream needs to be flagged. Not manually, not eventually. Immediately, automatically.
Truth requires decay. Knowledge goes stale. Last quarter's revenue number is a historical fact, not a current one. A system that treats all facts as equally current is a system that will eventually mislead you.
We built AttestDB because no existing database handles any of this. Relational databases store rows. Graph databases store edges. Vector databases store embeddings. None of them track whether what they're storing is still true, where it came from, or what to do when the facts change.
AttestDB stores claims.
A claim is a fact with a source, a confidence score, and a timestamp. It can be corroborated, contradicted, retracted, or decayed. It carries provenance as a structural property, not an annotation.
That one primitive — the claim — makes it possible to build AI systems that organizations can actually trust. Systems where you can trace any answer back to its sources. Where contradictions are surfaced, not hidden. Where stale data is flagged before it becomes a bad decision. Where one retraction cleans up everything downstream without a week of archaeology.
We believe every AI system will eventually need a truth layer. Not just memory. Not just context. Truth — verified, sourced, time-aware, and self-correcting.
AttestDB is that layer.