Skip to content

Skills and memory

Reusable instructions the agent applies as building blocks. Yours can be created, edited and deleted; built-in ones are read-only.

A skill is a SKILL.md file (frontmatter plus a Markdown body):

  • slug — lowercase Latin letters, digits, hyphen and underscore, up to 80 characters; fixed once created;
  • category — custom, imported, auto-generated, anti-pattern; also fixed;
  • description — what it does and when to apply it, in one line;
  • content — step-by-step instructions;
  • apply: Always — a style or policy in force on every answer; When relevant — the model applies it by the meaning of the request.

Each skill shows its run count. Deleting is irreversible and confirmed.

A skill can be put on a schedule — with a rule using the “Custom scenario (skill)” analyser.

A knowledge base: articles on topics and entities assembled from your data.

A separate screen shows how much the agent has accumulated:

Layer What it shows
Wiki number of facts
Conversation memory messages held, how many are stale, size
Links edges and concepts
RAG chunks chunk count and index size
Storage knowledge, database and cache size

A layer expands so you can inspect its contents. If memory is empty, the screen explains that it fills up by itself as you work with notes, mail and tasks.

The agent only sees what is permitted. A source switched off in Settings → Data and privacy isn’t read — and it will say it found nothing rather than invent something.

About the app itself the agent answers from the built-in help. If it claims a feature doesn’t exist, the article may be missing, not the feature.

Background work is routed separately from chat — see Models, router and privacy.