Skills Are Procedural Memory

Hermes Agent's learning loop needs a container for more than facts. This chapter explains that container: a skill is a reusable procedure that loads when a class of task requires it, while memory holds compact facts and tools provide executable capabilities. The distinction helps a reader decide what should be saved after a successful interaction, how that saved guidance stays affordable in future prompts, and when an agent-created procedure is safe enough to reuse.

A skill is neither a memory entry nor a tool

A memory entry answers a broad question: what should Hermes know by default? It may record a stable project convention, a user's preferred communication style, or an environment fact that applies across sessions. Because memory is injected into new prompts, it must remain compact and broadly relevant.

A tool answers a different question: what operation can the runtime perform? A file reader, terminal backend, browser, or search integration has a handler, a schema, and an execution boundary. Tools can be called precisely, checked for availability, and wrapped in approval or isolation. A tool does not usually explain the reasoning sequence in which it belongs.

A skill answers: how should this class of work be approached? It can contain the order of steps, decision points, user-specific preferences, pitfalls, verification rules, templates, and supporting scripts. The skill tells the model when the workflow applies and how to coordinate available tools. It is procedural memory: a durable method, not a permanent fact and not a deterministic program.

That distinction gives Hermes's learning loop somewhere to put different kinds of experience. “The user prefers concise responses” is too broad for a task skill and too useful to leave buried in a transcript. “When reviewing a repository, inspect the tests before widening a helper signature” is too procedural for an always-on profile but valuable when the relevant workflow loads. “Run the test suite” is an action that belongs to a terminal tool. A good system keeps all three statements in their proper layers.

Skills are compatible with the agentskills.io open standard and use a structured document format. A main SKILL.md carries the trigger, procedure, pitfalls, and verification shape. Supporting material can hold topical references, templates, scripts, or assets. The point is not the folder structure itself; it is the separation between the always-relevant operating rule and depth that should be opened only when the task needs it.

Progressive disclosure keeps the library usable

If every installed skill were copied in full into every system prompt, procedural memory would become another form of context overload. Hermes instead presents a compact index first. The model can see names and descriptions, then load a matching skill, and then open a specific reference file if the workflow requires more depth. Full knowledge is therefore available without being permanently resident.

This is a routing mechanism as much as a storage mechanism. A description must be specific enough to signal when a skill matches, but the details belong in the body or references.