PROVENANCE

Answers you can prove

For agents that touch real decisions, 'the model said so' isn't enough. Agent Memory stores provenance as data on every fact, keeps a queryable trace of how each memory formed, and returns the exact sources behind every answer - so recall you can trust means recall you can defend.

Join the waitlist Technical deep-dive

01 |PROVABLE ANSWERS

Every answer points back to its source

Provenance isn't a bolt-on log. It's a stored field on every row and a queryable trace graph, so any answer can be walked back to the exact messages and records it came from - by the agent, or by an auditor - without a separate logging pipeline to reconcile.

Built to be defended

Provenance, audit, and access control are properties of the store itself - not application code you have to remember to write.

Provenance on every row

Where a fact came from is a stored field, not metadata bolted on afterwards.

Learn more

Queryable trace graph

How a memory formed is itself a graph you can traverse and audit.

Learn more

Cited answers

Recall returns the exact records behind each result, ready to show a stakeholder.

Learn more

Nothing overwritten

Superseded facts are kept, so you can always see what changed and when.

Learn more

Deny by default

Per-verb grants govern who can read and write which memories.

Learn more

Defensible by design

When an answer reaches a decision-maker, the trail behind it is already there.

Learn more

02 |ONE SUBSTRATE

Trust is a storage property, not a promise

Middleware bolts provenance on with external logging and hopes the two stay in sync. Agent Memory makes provenance part of every write, in the same ACID transaction as the fact - so the trail can never drift from the data.