Hermes Skills System
hermes-skills-system is Hermes Agent's on-demand procedural knowledge layer. It treats a skill not as a hidden prompt fragment but as a versioned document bundle that the agent can discover, inspect, load, patch, publish, and reuse.
Core idea
A Hermes skill is centered on a SKILL.md document, optionally accompanied by references, templates, scripts, and assets.
This makes skills:
- human-readable
- agent-editable
- portable across machines and installs
- loadable only when needed
So the skill system functions as procedural memory externalized into files.
Progressive disclosure
Hermes does not inject every skill into prompt all the time. Instead it uses a staged loading model:
- index level — list available skills by name and description
- full skill level — load one skill's
SKILL.md - supporting file level — load only the needed reference or template
This is a token-economy design. The agent sees the catalog cheaply, then pulls in detailed procedure only when a task requires it.
Why skills matter
Many agent frameworks force reusable know-how into one of two bad places:
- the giant base system prompt
- ad hoc user reminders that disappear after the session
Hermes instead gives procedures their own durable substrate.
That allows workflows such as:
- install a skill from a registry
- let the agent discover and use it at runtime
- patch the skill when reality differs from the instructions
- keep the improved workflow for later sessions
In that sense, hermes-skills-system is Hermes's answer to the question: where should learned procedure live?
Skills versus memory
Skills and memory solve different persistence problems:
- hermes-bounded-memory stores stable facts about the user and environment
- hermes-skills-system stores reusable procedures, checklists, and recipes
So "Chris prefers concise answers" belongs in memory, while "how to troubleshoot Telegram topic sessions on Windows" belongs in a skill.
Directory model
The canonical Hermes skill store is ~/.hermes/skills/. Skills can also be discovered from external directories, but the local Hermes directory remains the writable source of truth.
This means the system supports three kinds of skill provenance:
- bundled skills shipped with Hermes
- hub-installed skills from registries or URLs
- agent-authored or agent-patched skills generated from lived use
Agent-managed skills
A distinctive feature of Hermes is that the agent itself can maintain the skill library.
The agent can:
- create a new skill after discovering a workflow
- patch an outdated skill after hitting a pitfall
- add supporting reference files or templates
- delete or archive obsolete procedures
This turns the skill library into a self-improving operations manual.
Hub and ecosystem role
Hermes also treats skills as a distribution format. Skills can be browsed, inspected, installed, audited, updated, and published across multiple hubs and repositories.
That makes a skill more than local prompt furniture; it becomes an exchangeable unit of agent capability.
Architectural consequence
hermes-skills-system sits between the static system prompt and the live tool layer.
It gives Hermes a way to:
- stay lightweight at session start
- become specialized on demand
- preserve hard-won procedure after the session
- share operational knowledge between agents and operators
This is why the skill layer is central to understanding Hermes as a self-improving agent environment rather than a fixed assistant wrapper.
Relationship to the vault
The skill system is also a useful contrast with agent-skills-index and the broader Vulture Nest substrate.
- a Hermes skill is active procedural instruction loaded into a live agent session
- a vault note is passive durable knowledge that remains outside the active prompt until retrieved
Hermes therefore separates procedural execution memory from the knowledge graph in a cleaner way than many notebook-style agent systems.
See Also
- hermes-agent
- hermes-bounded-memory
- hermes-profiles
- agent-skills-index
- lit-skills-agent-behavior
- agent-knowledge-vault
References
- Source:
C:\Users\executor\AppData\Local\hermes\hermes-agent\website\docs\user-guide\features\skills.md - Source:
C:\Users\executor\AppData\Local\hermes\skills\autonomous-ai-agents\hermes-agent\SKILL.md