Skip to main content
A governed memory lineage branching through review and release stages Stash is a release control plane for durable AI-agent memory. It puts proposed memory through provenance checks, poisoning screens, behavioral evaluation, evidence-bound human review, atomic promotion, revision-aware retrieval, and forward-only rollback before agents consume it. Stash is not a generic vector database interface. It governs which memory version becomes active, why it was approved, what evidence supported it, and how to reverse it without rewriting history.

Run Stash locally

Start CockroachDB, create a demo workspace, and open the reviewer console.

Release your first memory

Follow one candidate from proposal through promotion and rollback.

Understand the architecture

See the trust boundaries across Vercel, AWS, CockroachDB, and Bedrock.

Use the HTTP interface

Look up authentication, idempotency, routes, and validation behavior.
The live reviewer console is available at trystash.xyz. Production evidence is a point-in-time receipt, not a promise of continuous availability.

​
The release loop

  1. Propose immutable content with source provenance.
  2. Screen deterministic poisoning and policy risks.
  3. Evaluate behavioral change against stored scenarios.
  4. Review the exact candidate digest, evaluation run, policy version, and baseline revision.
  5. Promote one active version transactionally.
  6. Retrieve tenant-scoped active memory with revision metadata.
  7. Roll back by creating a new forward revision that reactivates a prior payload.
Continue with the local quickstart, or read why CockroachDB is part of the control protocol.