Skills & Procedural Memory

If memory is what the agent knows, skills are how it does things — procedural memory. A skill is a SKILL.md plus its bundle, surfaced through progressive disclosure so the agent loads detail only when it needs it. The practical edge here is the learning loop: letting Hermes create and refine skills from experience, then curating that so the library stays sharp.

Treat the learning loop as a feature to exploit, not a thesis to write. Let the agent draft skills from what it just did — then prune. An uncurated skill library fills with vague, half-overlapping procedures and quietly gets worse.

Planned notes

• Skills as procedural memory (vs factual memory)
• The SKILL.md format and a minimal example
• Progressive disclosure: loading skill detail on demand
• Skill bundles: scripts, templates, and assets alongside the instructions
• The open skills standard (agentskills.io / Skills Hub)
• The learning loop: agent-authored and self-refined skills
• Curating generated skills so the library stays sharp
• Verifiable skills: a skill that can check its own output
• Skill vs tool vs MCP vs profile — which abstraction fits
• Starter skills: research, applications, proposals, atlas maintenance, PR review, inbox triage

Core sources

• Hermes — Feature: Skills — https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
• Skills Hub — the open skills standard. https://agentskills.io
• Hermes — Features overview — https://hermes-agent.nousresearch.com/docs/user-guide/features/
• Hermes Agent — repository (self-improving skills) — https://github.com/NousResearch/hermes-agent

Connects to: Tools, MCP & Plugins · Profiles, Actors & Delegation · Native Orchestration · Income Workflows