VERIFIABLE MEMORY

Verifiable memory
for AI agents

An answer is a moment. Evidence gives it a history. PHIROK keeps the source, context and policy behind an AI decision together, so the past can be inspected instead of reconstructed from memory.

THE PROBLEM

An answer alone is not enough

When an AI makes a consequential decision — escalate a case, approve an exception, change a record — the answer alone does not explain itself. What source did it use? Which policy was active? What context was available at that moment?

Chat logs and dashboards can be edited, reordered or lost. Reconstructing the past from them is guesswork, and guesswork is a poor basis for trust in production systems.

WHAT PHIROK ADDS

One record, a complete chain of context

PHIROK records evidence around each decision, so the moment can be revisited as it was recorded:

  • Evidence records that bind source, context, policy and a receipt reference together.
  • Append-only history: records are added, not rewritten, so the past stays as it was recorded.
  • BLAKE3 hash-linked lineage: changes to protected bytes are detectable against a trusted reference.
  • Tenant-scoped records: evidence stays inside its workspace.
  • Deterministic replay of stored state: return to the state as recorded, not a retelling of it.

HOW IT IS USED

Evaluation first, then the integration contract

PHIROK is evaluated through an evaluation brief for a consequential AI workflow. The integration contract covers tenant scope, centralized authentication, authorization designed to fail closed, and authoritative receipt handling.

It is a layer for systems where decisions carry consequences — not a chatbot feature and not a claim about how well a model reasons.

CAPABILITIES AND LIMITS

What this proves — and what it does not

A hash check detects whether bytes changed against a trusted reference. It does not prove the content is true, and a receipt is a hash-linked reference, not a digital signature.

Replay returns stored memory state; it does not guarantee identical model output.

Hash-linked storage, tenant boundaries and deterministic replay address specific problems. They do not eliminate hardware failures, invalid input, scheduling bugs, or the need for backups and recovery procedures.

A VERIFIED EXAMPLE

Touch the evidence in the playground

The evidence playground on this site shows sample decision records you can inspect, seal, alter and re-check in your browser. It is a demonstration: it computes SHA-256 hashes locally over sample text and is independent of PHIROK's production BLAKE3 receipt implementation. Sample references are not live receipts.

See the playground, then read how the platform and the trust model describe the same mechanism.

RELATED GUIDES

Continue with another guide

See how the evidence layer works

The platform page describes the six capabilities built around one principle: memory should remain inspectable.

Explore the platform

YOUR NEXT DECISION

Build with memory.
Stand behind the evidence.

Give consequential AI a history you can inspect.