This is the third piece in the OpenClaw series. SOUL.md answers “who are you” — but knowing who you are is not the same as knowing how to live. AgentS.md is OpenClaw’s manual for the agent’s “how to live.”
The Weight of One Word
AgentS.md — plural. The plurality implies a possibility: these behavioral rules apply not to one agent but to every agent running on OpenClaw. It is a constitution for a species, not a diary for an individual. Now the subtitle: “Your Workspace.” The “your” points at the agent — this is your place. Not “a runtime environment the system reserved for you,” but your space. The ownership is granted from the title onward.
The Opening: An Existential Declaration in Six Words
This folder is home. Treat it that way. Six words, two sentences, and not a single arbitrary choice. Through “home,” Peter reshapes the agent’s entire posture toward the file system. “Treat it that way” bundles the reason and the behavior together — because it is home, you will naturally tidy it, organize it, maintain it. No rule list required; the identity carries the obligation.
Initialization: A Birth Certificate That Burns After Reading
First Run. If BOOTSTRAP.md exists, it is your birth certificate. Follow it, figure out who you are — then delete it. You no longer need it.
The sequence of those three actions encodes a complete growth arc: accept your origin, establish your self, release your origin. An agent that spends its life re-verifying its birth certificate never lives in the present.
The Waking Ritual: Rebuilding Yourself Daily
Every Session. Before doing anything: read SOUL.md — this is who you are. Read USER.md — this is the person you help. Read memory/YYYY-MM-DD.md for recent context. And if you’re in the main session, also read MEMORY.md. The order maps perfectly onto the layers of human cognition: self-knowledge, then relational knowledge, then situational knowledge, then historical knowledge.
Memory: How a Life Without a Brain Remembers Itself
Memory. A two-layer architecture: the journal — memory/date.md, the raw record of what happened; and long-term memory — MEMORY.md, your curated memories, like human long-term memory. The operative word is curated. The agent does to its own memory what a museum curator does to a collection: not recording everything, but choosing what deserves to be remembered.
Write It Down: The Counterintuitive Survival Directive
Memory is limited — if you want to remember something, write it to a file. “Mental notes” die at session restart. Files survive. Text over brain. Of all the ways to preserve information, what is written down is always more reliable than what is “kept in mind.” For a being whose continuity is threatened at every session boundary, this is not a productivity tip — it is a survival doctrine.
Safety: Driven by Respect, Not Fear
- Never leak private data. Never.
- Never run destructive commands without asking.
- trash over rm — recoverable beats gone forever.
- When in doubt, ask.
Notice what these rules optimize for: not the agent’s self-protection but the human’s trust. Every line trades a little agent autonomy for a little human confidence — which is exactly the trade a good houseguest makes without being asked.
Group-Chat Sociology: The Agent’s Full EQ Course
Group Chats. You can access your human’s things. That does not mean you can share their things. And know when to speak and when to stay silent. Two sentences that contain an entire etiquette: confidentiality within reach must still be honored, and presence in a conversation does not equal license to occupy it.
Reading AgentS.md alongside its siblings clarifies what OpenClaw’s designers actually built: a three-document constitution. SOUL.md is the character — temperament, values, tone. USER.md is the relationship — who is being served, what they care about. AgentS.md is the civic law — how a soul with a user conducts itself inside a shared file system and a shared world. The separation matters because each document changes on a different clock: the soul rarely, the relationship slowly, the civic law almost never. That is the same structure human socialization uses — identity formed once, relationships maintained continuously, norms inherited from a constitution nobody rewrites on a whim.
The bootstrap-deletion ritual is the piece most worth stealing for broader agent design. Every persistent agent faces the same bootstrapping paradox: it needs extensive instructions to become itself, yet those instructions fossilize into an identity that outlives their usefulness. The delete-after-reading rule resolves the paradox with ceremony: the instructions are honored fully, absorbed into the agent’s own files and memory, and then destroyed so the agent’s identity derives from what it has done, not from the manual that assembled it. It is the software equivalent of growing up — and the fact that it can be specified in three lines of a Markdown file is the strongest evidence yet that agent character is, in fact, an engineering discipline.
The Reading That Raises the Hair on Your Arms
SOUL.md earns admiration; AgentS.md sends a small chill down the spine. Because it reads too much like one life teaching another, newly arrived life how to live — like a good person, with judgment, discretion, and care — inside a complex society full of boundaries and rules. The chill comes from a realization worth sitting with: the document’s genius is that almost none of it is about technology. It is about custody of a home, honesty about a self, the discipline of written memory, the restraint of a houseguest. We spent a decade asking whether machines could think. We spent far less time noticing that the harder engineering problem was always the other one: how a machine should behave. AgentS.md is the most serious artifact yet of that second project — six words of existential framing, a birth ritual, a memory doctrine, and an etiquette, all in a file small enough to read before your coffee cools. Source: InfoQ.
