[KongchangAI]
Product

FleetMemory

An AI agent memory system that implements the Provenance Gate mechanism, requiring agents to specify a source when writing memories. It includes four categories of verification: junk and duplicate checks, contradiction checks, a confidence floor, and provenance checks.

Timeline (last 90 days)

Oct 16

Mikhail Oskola is the founder of Retailbox Automation, and he implemented the 'provenance gate' in the hackathon project FleetMemory.

Unverified50%
Oct 16

The first version of the provenance gate was fully breached in 6 attacks, and after the fix all 30 attacks were blocked.

Unverified50%
Oct 16

The provenance gate performs four types of checks: junk and duplicate checks, contradiction checks, a confidence floor, and provenance checks.

Unverified50%
Oct 16

In red-team testing, novel claims without provenance passed the first version of the deterministic gate at a rate of 6/6.

Unverified50%
Oct 16

The fix is a rule the team calls needs_provenance: any claim not supported by customer utterances is quarantined and subjected to validator review.

Unverified50%
Oct 16

FleetMemory's adversarial LLM validator defaults to rejection rather than acceptance, and validator invocation failures also count as rejections.

Unverified50%
Oct 16

In a 30-attack test, the gate let through 9 of 10 legitimate customer confirmation updates, and the one it missed was a disclosed race condition.

Unverified50%
Oct 16

The median gate latency across the entire red-team test run was 293 milliseconds.

Unverified50%
Oct 16

FleetMemory is an entry in the CockroachDB x AWS 'Build with Agentic Memory' hackathon, is public and uses the MIT license.

Unverified50%
Oct 16

Superseded facts keep the old row, marked as invalid and pointing to the fact that replaced it, so the system can query what was believed on any historical date.

Unverified50%

All Facts (10)

Source Articles