SOUL.md
Who the agent is.
Identity, voice, values, hard limits.
Plain markdown you can save, edit, and version. Download one file or all seven, drop them into a folder, and point your agent at them. Same sources as the interactive overview. For a one-page map of roles and read order, open README.md (also under /templates/).
Tip: if Download is blocked, use Open raw then “Save as…” — or Copy to paste straight into your editor.
Who the agent is.
Identity, voice, values, hard limits.
Where the agent lives.
Metadata, routing, model config, multi-agent settings.
How the agent operates.
Workflows, decision rules, escalation paths, communication patterns.
Who the agent serves.
You. Preferences, projects, contacts, flagged priorities.
What the agent has learned.
Evergreen long-term memory. One fact per line. Never deleted, only superseded.
How the agent acts on its own.
Scheduled tasks. Morning briefs. Weekly reviews. Conditional alerts.
What the agent can reach for.
Every skill, MCP, and built-in capability — with usage rules.
Use the cards above to grab each file, or the ZIP / curl block below for all seven at once. To generate filled files from a questionnaire, use Build instead of editing blanks by hand.
Grab every file in a single archive — or stream them with curl when you already know your origin.
Replace YOUR_ORIGIN with your deployed site (e.g. https://example.com).
# Download all seven into the current directory BASE="YOUR_ORIGIN" for n in SOUL IDENTITY AGENTS USER MEMORY HEARTBEAT TOOLS; do curl -fsSL "$BASE/templates/${n}.md" -o "${n}.md" done
Same seven markdown files — different hosts. Copy the folder, point your runtime at it, version in git.
The folder your gateway loads each session
Point OpenClaw at a directory of markdown. These seven files are that directory — no proprietary schema. Add a memory/ subfolder for daily notes; the spec already describes the split.
Personality, values, hard limits — core context on boot
Model, routing, multi-agent metadata
Session checklist, workflows, decision table
Who you are, priorities, “do not do”
Evergreen facts (promote from memory/ daily files)
Scheduled tasks + autonomy boundaries
Skills, MCPs, selection order
Same markdown, different host
In Cursor you often merge SOUL + AGENTS into .cursor/rules or AGENTS.md in the repo. TOOLS.md documents MCP servers; USER.md is optional context. You are not locked in — export or symlink the same files.
Tone and boundaries → rules or always-on context
How the agent decides and escalates
Which MCPs exist and when to use them
Preferences loaded when pair-programming
Model choice hints; less critical in IDE
Capabilities without losing the soul
TOOLS.md is the static map of what the model may call. AGENTS.md defines selection order and failure handling. HEARTBEAT.md covers unattended checks — never confuse “can call API” with “may act without you”.
Every MCP, skill, built-in — parameters & errors
Tool selection guide + escalation when tools fail
Rate limits, health checks, no silent destructive acts
Confirmed facts before hitting external APIs again
Hard limits on what tools must never do
Export a Cursor rules bundle from Build after filling the questionnaire.
Every edit to templates/*.md should get an entry under content/changelog/ in the same commit.
Added personality depth, core values, and explicit failure-mode behaviour.
First public starter template — voice, values, capabilities, and hard limits.
First public starter template — metadata, routing, and model configuration.
Added multi-agent and session settings blocks for routing metadata.
First public starter template — session checklist, workflows, and decision rules.
Expanded six-phase pipeline pseudocode, decision table, and escalation paths.
First public starter template — preferences, projects, and standing “do not do”.
Expanded contacts, recurring tasks, and flagged priorities sections.
Documented memory hygiene — one fact per line, supersede instead of delete.
First public starter template — evergreen facts, hygiene, and supersede rules.
Scheduled monitoring tasks with explicit “no destructive acts without confirmation”.
First public starter template — scheduled tasks and autonomy limits.
First public starter template — capabilities map and tool selection order.
Clarified tool selection order, error handling, and unavailable-tool stubs.
Personal learning project — not a service SLA. Open a GitHub Issue when something is wrong or unclear.
templates/ plus a content/changelog/ entry in the same PR.