MP mikepatraw.com

Hermes Agent System / Case Study

Skill Curator Audit for Agent Memory Hygiene

A non-destructive audit loop for keeping assistant skills useful, current, and conflict-free.

Draft public-safe Public-safe portfolio draft

Executive Summary

An assistant that learns over time needs governance. This case study shows a report-first curator pattern that improves memory hygiene without unsafe automatic deletion.

Problem

What needed solving

An assistant that learns over time can accumulate stale, duplicate, or conflicting skills. Without maintenance, the learning loop becomes a liability instead of an advantage.

Context

Why this mattered

Several indexed Hermes/Curator videos emphasized skill maintenance. Fred already had a growing skill library, so the useful next step was a local audit that reports problems without making unsafe automatic edits.

Constraints

Operating boundaries

  • Never delete skills automatically.
  • Require approval before destructive or broad skill changes.
  • Run locally without external services or model calls.
  • Prefer report-only evidence over silent mutation.

Build

What I built

  • A report-only auditor for local skills.
  • Tests for audit behavior.
  • Operating guidance that distinguishes safe maintenance from risky auto-pruning.

Architecture

Workflow

  • Input: local skill files and metadata across Fred’s profile.
  • Analysis: identify stale, duplicate, overlapping, or potentially conflicting skill procedures.
  • Output: audit report for review, with no automatic deletes.
  • Control: patch procedures only after a concrete, evidence-backed change.

Security / Operations

Key decisions

  • Non-destructive by default because skills encode procedural memory and mistakes can compound across sessions.
  • Audit results are treated as triage, not authority.
  • The Curator concept was adapted to Fred’s stronger safety boundary: report first, modify only when justified.

Impact

What changed

  • Improved long-term assistant reliability by adding maintenance pressure without hidden deletion risk.
  • Captured a practical memory-hygiene story for the portfolio: learning systems need governance, not just accumulation.
  • Reduced the odds that outdated skills override newer operating procedures.

Evidence

Source trail

gemini-minimal curator · skill

Hermes Agent Curator Guide: FIX Your Agent Skills!

curatorskillstalemaintenance
View source
gemini-minimal curator · skill

Hermes Agent V0.12 Just Changed AI Agents Forever!

curatorskillskills
View source
gemini-minimal curator · archive

Hermes Agent Just Got 10x BETTER (Curator Explained)

curatorarchiveskills
View source

Next

What I’d improve

  • Add richer duplicate detection by embedding or structural similarity if local-only methods are not enough.
  • Track skill usage timestamps more formally.
  • Surface audit summaries in periodic PA systems audits.

Public-safe notes

Publishing boundary

  • The public story is memory hygiene and governance, not internal skill contents.
  • No automatic deletion is presented as a feature, not a limitation.
  • The case study connects applied AI capability with operational risk control.