Hermes Agent /learn feature goes live: Local files, web pages, and conversations become Skills with one click

Nous Research launches the /learn feature for the open-source AI Agent Hermes, allowing the Agent to autonomously gather materials, generate skill files, and store them in the skill library, turning "once-used operation workflows" into reusable callable tools without requiring manual organization by engineers.

(Previous Highlight: AI Renaissance — Philosophers Become Hot Commodity in AI Labs, Embedding Ethics into Your Model)

(Background Supplement: US-China AI Rivalry Intensifies, but Scholars from Both Countries Agree: Don't Let AI Have Its 'Chernobyl Moment')

Nous Research has added the /learn feature to the Skills system of its open-source AI Agent "Hermes". The operational logic is straightforward: you tell the Agent what you want to learn — a local SDK folder, an online documentation page, or the deployment process you just walked it through — and it uses its existing tools to autonomously gather materials, then produces a properly formatted skill file, stored in ~/.hermes/skills/, ready to be called next time without additional tools.

Turning 'What You Just Did' into a Reusable Tool for Next Time

The official positioning of /learn is: to quickly turn 'what you already know' or 'a bunch of reference materials' into a reusable skill, without having to write a SKILL.md manually.

It supports four types of material sources, each with corresponding typical scenarios:

The first is a local library or SDK folder. Example command: /learn the REST client in ~/projects/acme-sdk, focus on auth + pagination. Suitable for solidifying operational knowledge of internal team tools into a skill;

The second is an online documentation page. Example command: /learn https://docs.example.com/api/v2. Suitable for quickly ingesting third-party API documentation, saving the cost of repeated lookups;

The third is the complete workflow you just walked the Agent through in the conversation. Example command: /learn how I just deployed the staging server. It condenses a one-time operation into reusable steps for next time;

The fourth is any pasted verbal notes or unstructured text. Anything you can describe can theoretically be fed into it, with almost no boundary on openness.

After receiving the request, the Agent uses its existing tools — read_file, search_files, web_extract — to autonomously gather materials, then generates a skill according to built-in writing specifications: description limited to 60 characters, fixed chapter order, use of Hermes tool terminology, no self-invented commands.

Self-improvement becomes more concrete

Traditional approach: An engineer observes the Agent complete a task, manually or lets AI organize it into a documentation file, then writes that file into a skill, so the Agent can use it next time. In the entire chain, the human is the sole 'knowledge extractor'.

The capability boundaries of most AI Agents are either hard-coded or rely on engineers manually updating prompts periodically. The skill library is static and does not automatically grow with usage.

/learn shortens this chain to: Agent completes a task → user issues a command '/learn that workflow just now' → skill generation complete. The human steps back from being the 'knowledge extractor', leaving only the judgment of 'whether to learn'.

Hermes's skill library is dynamic, automatically expanding with accumulated usage scenarios. However, a reminder here: just because the capability library can grow doesn't mean everything that grows is correct; quality optimization is still necessary.

Hermes was originally positioned as a 'self-improving agent' - not only completing tasks but also remembering and accumulating reusable skills, getting smarter with use. /learn turns this positioning from concept to concrete operation: the agent not only executes but can also precipitate one-time operations into reusable assets.