HARNESS MEMORY

Memory for agent
harnesses

Agent = model + harness. The model reasons; the harness is everything around it that lets it act — tools, loop, guardrails, and memory. When the harness has no durable memory, multi-step work collapses the moment a session ends.

THE PROBLEM

The harness is where continuity lives

A harness orchestrates: it runs tools, carries context, enforces guardrails and recovers from failure. Most of that state lives in session files or logs that can be edited, lost or silently changed.

When something goes wrong, the question is not only “what did the agent do?” but “what did it know, from which source, under which policy — and can we prove it?” Without a verifiable memory layer, the harness has no trustworthy answer.

WHAT PHIROK ADDS

A verifiable memory layer inside the harness

PHIROK supplies the memory component of the harness as inspectable evidence:

  • Persistent records across sessions, bound to source, context, policy and receipt reference.
  • Append-only history, so the harness's past states remain available.
  • Tenant-scoped evidence, separated by workspace.
  • BLAKE3 hash-linked lineage for tamper evidence against a trusted reference.
  • Replay of stored state, so a past moment can be revisited as recorded.

HOW IT IS USED

Through the integration contract

The product architecture includes protected MCP and gRPC boundaries, with authentication, tenant scope and authoritative receipt handling as part of the integration contract, and authorization designed to fail closed. Evaluation starts with an evaluation brief for a consequential workflow.

CAPABILITIES AND LIMITS

A layer, not a harness — and no framework claims

PHIROK is a memory and evidence layer; it is not itself a harness, and this page claims no compatibility with any specific harness framework or agent. A hash check detects change against a trusted reference — it does not prove content true — and a receipt is a hash-linked reference, not a digital signature.

Replay returns stored state, not identical model output.

A VERIFIED EXAMPLE

Sample records you can inspect

The evidence playground demonstrates the evidence model with sample decision records and a local hash check in your browser. Samples are labeled; the browser check uses SHA-256 for demonstration and is independent of the production BLAKE3 implementation.

RELATED GUIDES

Continue with another guide

Read the architecture

The architecture page follows authenticated evidence capture into append-only storage, lineage, retrieval and review.

Explore the architecture

YOUR NEXT DECISION

Build with memory.
Stand behind the evidence.

Give consequential AI a history you can inspect.