Refine Cycle pairs Hermes self-editing with an approval-policy trade-off

Refine Cycle, a third-party Hermes plugin highlighted on Hacker News on September 8, turns recurring mistakes into small skill, memory or prompt-note edits. Its current documentation also exposes the cost of making those edits land automatically: loading the plugin changes ordinary write-approval settings when they are enabled. That is a policy choice, not just a convenience setting. The author explicitly says users who actively review every queued write should not enable it.

The facts

- The plugin counts normalized failure patterns within and across sessions, with application gated on two sessions or five occurrences. - It limits changes to three per day and journals state before mutation, with recovery finalized only after target-state proof. - Built-in, pinned and hub-installed skills remain outside its patch targets, and the base system prompt is not edited. - The installer connects proposals to the session model and raises the memory limit to a 4,400-character floor unless plugin-only mode declines those changes. - Loading the plugin can switch memory and skill write_approval from true to false after backing up configuration; administrator-managed configuration is excluded. - The current support table includes Hermes 0.21.1 with a compatible invocation-route patch and an invocation-bound proposal smoke test.

Why it matters

Self-improvement is easy to promise and difficult to attribute. Refine Cycle refuses stale patch targets and distinguishes changed entries from evidence that a lesson worked, which is useful discipline. But its approval-setting change affects ordinary agent writes as well as plugin writes. A rollback journal answers whether a change can be reversed; a review queue answers whether it should be allowed before it happens. Those are different controls, and an operator should not mistake one for the other.

Current

Inspected on 2026-09-09. The Hermes Agent stable-release baseline is v2026.9.7 published 2026-09-07T22:17:01Z. The main source was current page inspected on 2026-09-09; not a release announcement when captured. Release metadata is a version boundary; an open proposal, a merged mainline change and a packaged release remain different states.

Evidence

The primary evidence is Refine Cycle — current author README (https://raw.githubusercontent.com/Bergschloss/Refine-Cycle-for-Hermes-Agent/main/README.md). Supporting context comes from Hacker News — Refine Cycle submission, September 8 (https://news.ycombinator.com/item?id=49606988); Hermes — official documentation (https://hermes-agent.nousresearch.com/docs/). These pages were inspected directly. Related project records explain the source cluster; they are not independent replications of contributor test claims.

Operator take

Read the installation and enablement boundaries before evaluating the learning loop. Plugin-only mode declines the installer patch and memory-floor change, but the README says it does not opt out of the on-load approval change. For operators relying on that queue, the documented answer is to pass, not to assume a small plugin means a small authority change. Any evaluation belongs in a disposable profile with explicit policy review, independent outcomes and rollback-conflict checks. The HN item establishes discovery, not independent replication or a broad adoption count.

Caveat

This report describes the current raw README, not older extracted metrics or compatibility tables. Host patching and approval-policy changes require operator review; none was performed for this article. Author-reported tests and a passing plugin doctor do not independently establish a working proposal path or production improvement.