openclawbrain-site

SOP: Create a new OpenClaw agent + dedicated OpenClawBrain profile (TypeScript-first)

Canonical flow for a new profile-style setup:

• one OpenClaw agent workspace
• one dedicated OpenClaw runtime brain profile
• one pinned OpenClawBrain package set
• explicit routing/binding in OpenClaw runtime

A) Decide IDs and paths

agentId="pelican"
agentName="Pelican"
workspaceDir="$HOME/.openclaw/workspace-pelican"

Conventions:

• agentId is lowercase and stable (main, pelican, bountiful, …)
• each agent has its own workspace
• each agent has its own OpenClaw brain profile

B) Create workspace skeleton

mkdir -p "$workspaceDir/memory"

: > "$workspaceDir/AGENTS.md"
: > "$workspaceDir/SOUL.md"
: > "$workspaceDir/USER.md"
: > "$workspaceDir/MEMORY.md"
: > "$workspaceDir/IDENTITY.md"
: > "$workspaceDir/TOOLS.md"
: > "$workspaceDir/active-tasks.md"
: > "$workspaceDir/memory/$(date +%F).md"

Populate policy/persona files before live traffic.

C) Build and verify OpenClawBrain package set

Run from the OpenClawBrain TypeScript workspace root:

corepack enable
pnpm install
pnpm check
pnpm release:pack

Promote one versioned pack set for this rollout.

D) Add agent brain profile in OpenClaw runtime

Add a dedicated profile for this agentId with defaults on:

{
  "agentId": "pelican",
  "brain": {
    "enabled": true,
    "packSet": "<versioned-pack-set>",
    "fastBootFromExistingFiles": true,
    "backgroundLearning": {
      "enabled": true,
      "prioritizeNewEvents": true
    },
    "labels": {
      "human": true,
      "self": true
    },
    "scanner": {
      "enabled": true
    },
    "harvest": {
      "enabled": true
    },
    "continuousGraphLearning": {
      "enabled": true,
      "decay": true,
      "hebbianCofiring": true,
      "structuralUpdates": true
    },
    "teacher": {
      "onHotPath": false
    }
  }
}

E) Bind routing

Bind inbound channels to the new agentId in OpenClaw runtime config.

Minimum checks:

• new account/channel maps to agentId
• no overlap with existing agent bindings
• fail-open behavior retained for this route

F) Add workspace policy snippet

Copy into SOUL.md:

## Always-on memory policy (default)

- Treat clear user corrections as immediate labels.
- Treat durable user teachings as first-class labels.
- Keep human + self labels enabled.
- Keep scanner and harvest enabled.
- Keep teacher off the hot path.
- Never store secrets in workspace memory files.

G) Verification checklist

Runtime checks:

• route reaches agentId
• first response works immediately from existing files
• background learning workers are active
• new events are prioritized over backlog
• labels/scanner/harvest loops are active
• continuous graph learning updates are visible

Boundary checks:

• OpenClaw owns serving/runtime/fail-open
• OpenClawBrain owns contracts, event/export normalization, provenance, packs, activation, compiler, learner

Proof checks:

• mechanism proof collected (contracts/provenance/normalization)
• product proof measured separately (quality/reliability outcomes)

H) Common footguns

• sharing one workspace across multiple agents
• deploying mixed package revisions instead of one pack set
• disabling labels/scanner/harvest “temporarily” and forgetting to re-enable
• putting teacher logic on hot path
• claiming product gains from mechanism checks only