ADR-0003: The reviewer's memory and state are Meterless's H-MEM and World Model, built in Go and checked by Meterless's own conformance suites

Status: accepted · Date: 2026-10-17 Decided by: the owner · Written by: the advisor (claude-code, claude-opus-5-5), 2026-10-18 Follows: ADR-0001 (Go) Trigger: Meterless is required in every entry (planning/CONVERSATION_2026-10-17.md, hops 11 and 12)

Context

Meterless is not a hosted service. It publishes open specifications of agent engines (Apache-2.0), each with a TypeScript reference and a conformance suite, meant to be handed to a coding agent. The ones with code today:

• H-MEM: hierarchical memory as a knowledge graph.
• World Model: user, task and environment state, event-sourced.
• Markovian: reasoning compression for long tasks.

A reviewer without memory repeats itself. A finding that was refuted in one pull request comes back in the next one, and a maintainer's ruling ("we do not validate here, on purpose") is forgotten.

The hackathon wants logs that show what each agent did, handoffs, critique and retries, and a trace that is not hand-written.

proven is written in Go (ADR-0001). The H-MEM conformance suite loads the implementation as a TypeScript module (HMEM_IMPL=<path to index.ts> npx tsx conformance/runner.ts).

Decision

1. H-MEM is the reviewer's memory across pull requests: refuted findings, maintainers' rulings, and the evidence behind both.
2. World Model is the state of one review, as a stream of events: every claim, piece of evidence, refutation and verdict. The trace in logs/ is that stream, not text written next to it.
3. Both are built in Go from the Meterless specifications, by our Cursor executors.
4. Both are checked by Meterless's own conformance suites, through a thin TypeScript adapter that calls the Go code.
5. Markovian is not in scope for the build window. It is a candidate if the build leaves time.

Consequences

• We implement Meterless's engines; we do not copy their reference code. The specifications are declared as a pre-existing open-source component.
• The build machines and the image carry Node, only to run the conformance suites (train order proven-cursor-infra, A13).
• Not measured yet: whether a Go engine passes the suites through a TypeScript adapter. It is the first thing measured in the build. If it does not work, this ADR is revisited.
• The judges' trace is generated by the same mechanism that holds the review's state, so it cannot drift from what the agents did.