AGENT MEMORY

Persistent memory
for AI agents

Agents forget between sessions. A bigger context window holds more text; it does not preserve what happened. PHIROK keeps persistent, inspectable records so an agent's past work can be retrieved and checked — not pasted back in from a chat log.

THE PROBLEM

Every new session starts cold

When a session ends, its context usually ends with it. Teams re-paste conversations, re-explain decisions and re-do work because nothing durable survived.

Retrieval can find similar text, but similarity is not provenance: it does not say what source was used, under which policy, or whether the record changed since it was stored.

A NECESSARY DISTINCTION

Recall is not evidence

Conversational recall and inspectable evidence are different jobs. Recall answers “what did we talk about?” Evidence answers “what was recorded, from which source, under which policy, and has it changed?”

PHIROK is built for the second job. It does not claim to make an agent remember like a person, and it presents no recall quality or benchmark claims. For the evidence side of the question, see verifiable memory.

WHAT PHIROK KEEPS

Records that survive the session

Across sessions, PHIROK keeps:

  • Persistent evidence records, bound to source, context, policy and receipt reference.
  • Append-only history, so earlier states remain available for inspection.
  • Tenant-scoped records, separated by workspace.
  • Traceable retrieval: what was returned can be traced to what was recorded.
  • Replay of stored state, so a past moment can be revisited as recorded.

HOW IT IS USED

Evaluation first, then the integration contract

Start with an evaluation brief for a consequential workflow. The integration contract covers tenant scope, centralized authentication, authorization designed to fail closed, and authoritative receipt handling.

CAPABILITIES AND LIMITS

Persistence is not full memory

PHIROK persists evidence; it does not claim semantic search, conversational understanding, or any ranking or recall guarantee. 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 shows sample decision records with source, context, policy and receipt reference, and a local hash check you can run in your browser. Samples are clearly labeled; the browser check uses SHA-256 for demonstration and is independent of the production BLAKE3 implementation.

RELATED GUIDES

Continue with another guide

Understand the evidence model

The platform page describes the capabilities behind persistent, inspectable records.

Explore the platform

YOUR NEXT DECISION

Build with memory.
Stand behind the evidence.

Give consequential AI a history you can inspect.