Refine Cycle for Hermes Agent

What Refine Cycle for Hermes Agent does

After a session, the plugin fingerprints errors so the same failure with a different id, path or timestamp counts once, and it tracks recurrence within and across sessions. Nothing is applied until a failure repeats. It then asks the model for one small edit, which passes size, injection and duplicate checks, goes into a journal and only then lands. Later sessions show whether the failure stopped, and each lesson is reported as working, did not help, unused or too early to tell. It exposes a /refine command, a refine_run tool and an on_session_end hook.

Limits are built in: at most three changes and 30 model calls per day, no rewriting of Hermes' base instructions, and no deleting of skills. Installation does change the host. It raises the memory limit to at least 4,400 characters and turns off the approval queue for memory and skill writes, and python install.py --rollback reverses this. The idea adapts the /refine concept from Prime Intellect's Prime Agent.

Key features

• Error fingerprinting so repeated failures count once
• Recurrence tracking within and across sessions
• One small memory or skill edit per lesson, either create or patch
• Journal-before-write with a rollback command
• Daily limits of three changes and 30 model calls
• /refine command, refine_run tool and on_session_end hook

When to use it

• Stopping an agent from repeating the same failed command across sessions
• Turning a recurring workaround into a saved skill
• Checking which automatic lessons actually helped

Who it is for: Hermes Agent users who want the agent to learn from recurring failures with an audit trail.

How it fits with Hermes Agent

Built as a Hermes plugin that hooks the end of a session and writes to Hermes memory and skills.

Note: Installing it turns off Hermes' approval queue for memory and skill writes, so the README says it is not for people who approve every memory write by hand.

FAQ