hermes - 💡(How to fix) Fix [RFC] Skills System Architecture Redesign: From Static Files to Living Knowledge

Summary

This proposal argues for a fundamental redesign of the Hermes Skills system — from "skills as static files" to "skills as intelligent, evolving, collaborative knowledge units."

The current system has seven critical limitations that prevent Hermes from being a true learning agent system. This RFC proposes a comprehensive architecture to address them.

The Seven Problems

1. No Semantic Understanding — Skills can only be found by exact filename, not by user intent
2. No Deduplication — Multiple skills can do the same thing with no conflict detection
3. No Versioning/Evolution — Skills never improve from usage experience
4. No Relationships — Skills exist in isolation, cannot compose or reference each other
5. No Personalization — One-size-fits-all, doesn't adapt to user expertise
6. No Quality Metrics — No way to know if a skill actually works well
7. No Composition — Cannot combine skills to handle complex tasks

Industry comparison: Hermes is 2-3 generations behind LangChain, AutoGPT, and CrewAI in skills architecture.

Proposed Architecture

┌─────────────────────────────────────────────────────────────┐
│                  Skills Orchestration Layer                 │
│  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐       │
│  │ Intent  │  │ Context │  │ Skill   │  │ Compos- │       │
│  │ Parser  │→ │Analyzer │→ │ Selector│→ │ ition   │       │
│  └─────────┘  └─────────┘  └─────────┘  └─────────┘       │
├─────────────────────────────────────────────────────────────┤
│                    Skills Knowledge Layer                   │
│           Skills Knowledge Graph (nodes + edges)            │
├─────────────────────────────────────────────────────────────┤
│                    Skills Storage Layer                     │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐   │
│  │ Content  │  │ Metadata │  │ History  │  │ Metrics  │   │
│  │ Store    │  │ Store    │  │ Store    │  │ Store    │   │
│  │(files)   │  │(SQLite)  │  │(Git)     │  │(SQLite)  │   │
│  └──────────┘  └──────────┘  └──────────┘  └──────────┘   │
├─────────────────────────────────────────────────────────────┤
│                    Learning & Evolution Layer               │
│  Feedback → Pattern Analysis → Suggestions → Improvement    │
└─────────────────────────────────────────────────────────────┘

Six Core Components

1. Skills Knowledge Graph — Represent skill relationships (depends_on, alternative_to, complements)
2. Intent-Based Matching — Match user goals to skills without exact names
3. Composition Engine — Dynamically combine skills for complex tasks
4. Learning Engine — Collect feedback, analyze patterns, suggest improvements
5. Personalization Engine — Adapt skill presentation to user expertise
6. Quality Assurance — Validate, test, and score skill quality

Proposed SKILL.md Format

---
name: browser-automation
version: 1.0.0
category: automation
tags: [browser, cdp, playwright]

# New: Structured knowledge
ontology:
  concepts: [CDP, Playwright, headless]
capabilities:
  actions: [navigate, click, type, screenshot]
  preconditions: ["Chrome running"]
usage:
  when_to_use: ["web scraping", "testing"]
  when_not_to_use: ["simple HTTP requests"]

# New: Quality metrics (auto-updated)
metrics:
  success_rate: 0.92
  usage_count: 156

# New: Relationships
relationships:
  depends_on: [chrome-cdp-connection]
  alternatives: [puppeteer, selenium]
---

Implementation Plan (22 weeks)

Success Metrics

Why This Matters

This is not a nice-to-have improvement. This is a competitive necessity. The agent landscape is evolving rapidly, and skills architecture is becoming a key differentiator. Without this redesign, Hermes risks falling further behind.

The proposed implementation is incremental, low-risk, and backward-compatible. Old skills continue to work while new capabilities are added.

Full proposal document (27KB) available on request — includes detailed component designs, database schema, tool specifications, risk analysis, and technical references.

I'm willing to contribute to the implementation. Let's discuss!