Memory, skills and persona

Long-term memory recalled on every message, saved and searchable conversations, about 200 skills, and six persona files that shape how your agent talks and works.

#Memory

The agent keeps typed notes about you and your work (user, preference, decision, finding, blocker, note) and recalls the relevant ones into every message. It writes them itself with the memory tool when you share something worth keeping; you can manage them too:

smooth-agent memory list
smooth-agent memory add "I deploy with Docker on a 2-core VPS" --kind preference
smooth-agent memory search docker
smooth-agent memory pin 7          # always included
smooth-agent memory remove 7

Turn it off with [agent] memory = false.

#Conversations

Every conversation is saved in smooth.db and searchable, by you and by the agent (session_search). Long chats are summarised instead of cut off: beyond about 80% of context_tokens, older messages are condensed while the last keep_recent stay word for word.

smooth-agent sessions list
smooth-agent sessions search "invoice template"
smooth-agent run --resume          # continue the last session

#Skills

Skills are tested playbooks (a SKILL.md plus helper files) the agent reads when a task matches. About 200 ship with the agent, for coding agents, GitHub, documents, media, research, finance, productivity, smart home and more. The index is in the system prompt; the full text loads only when needed.

smooth-agent skills list                  # core skills
smooth-agent skills list --all            # + optional skills
smooth-agent skills search pdf
smooth-agent skills install <name>        # turn on an optional skill
smooth-agent skills install https://github.com/you/your-skill
smooth-agent skills remove <name>

Load order: workspace skills/ → ~/.smooth-agent/skills → extra skill_dirs → core → optional. The agent can write and improve its own skills (skill_manage); editing a bundled skill saves your own copy.

#Persona files

Six Markdown files in the workspace are part of the agent's instructions on every message:

They start empty and cost nothing until you (or the agent, when you ask) write in them. Each file is capped at about 6,000 characters.

Last updated: 2 November 2026