openclaw-workspace-structure

OpenClaw Workspace Structure

Explain and help the user navigate the standardized OpenClaw workspace layouts. Two levels exist:

1. System-level layout (~/.openclaw/) — Created by the OpenClaw CLI, holds global config and all agent workspaces
2. Per-agent workspace layout — The flatter pattern each agent uses inside its workspace directory

Per-Agent Workspace Layout

Each agent workspace follows this flatter layout with bootstrap files at root:

<agent-workspace>/
  AGENTS.md              # Operating manual (safety rules, delegation, memory workflow)
  SOUL.md                # Personality, boundaries, communication style
  IDENTITY.md            # Name, role, avatar, vibe
  USER.md                # Who the agent serves (private sessions only)
  MEMORY.md              # Long-term curated knowledge (200-line budget)
  TOOLS.md               # Available tools, integration notes
  HEARTBEAT.md           # Proactive monitoring checklist
  GROUP_MEMORY.md        # Info safe for group contexts (optional)
  agents/                # Sub-agent directories (team workspace)
    <agent-id>/
      SOUL.md
      MEMORY.md
      logs/
      output/
  skills/                # Workspace-specific skills
  memory/                # Daily logs
    YYYY-MM-DD.md        # Raw session notes
    archive/             # Compressed logs older than 14 days
  team-memory/           # Shared cross-agent knowledge
    market-brief.md
    icp-profile.md
    brand-voice.md
  config/                # Local configuration
    security.json
  output/                # Agent work products
  logs/                  # Operational logs

Memory details: See openclaw-memory skill for the four-layer memory model, bootstrap file loading order, daily log workflow, SCRIBE compression, shared memory patterns, and /oc-memory command.

System-Level Layout (CLI-Managed)

~/.openclaw/
  openclaw.json              # Central configuration
  models.json                # Model registry
  workspace/                 # Shared collaboration space
    CLAUDE.md                # Shared instructions for all agents
    TASKS.json               # Task tracking (RACI schema)
    SPRINT_CURRENT.json      # Active sprint data
    SHARED_KNOWLEDGE.json    # Semantic memory layer
    IMPROVEMENT_BACKLOG.json # Continuous improvement items (min 5)
    agents-workspaces/       # Per-agent isolation
      <agent-id>/
        IDENTITY.md          # Agent role definition
        SOUL.md              # Behavioral instructions
        TOOLS.md             # Available capabilities
        auth-profiles.json   # Credentials (chmod 600)
    comms/                   # Async messaging
      <agent-id>/
        inbox/
        outbox/
      broadcast.md           # System announcements
    processes/
      PROCESSES.json         # 23 process definitions
    standards/               # Coding, quality, docs, research standards
  skills/                    # Reusable SKILL.md definitions
  memory/
    lancedb/                 # Vector embeddings store

Agent IDs

Sixteen agents map to workspace directories:

Agent	ID	Role
Cooper	main	Primary orchestrator
Pixel	debugger	Debugging specialist
Vault	cybersecurity	Security operations
Sage	solution-architect	Architecture design
Oracle	predictive-analyst	Predictive analysis
Nova	nova	General purpose
Mirror	metacognition	Self-reflection
Vista	business-analyst	Business analysis
Forge	implementation	Implementation work
Vex	tester	Testing specialist
Axon	devops	DevOps operations
Vigil	quality-assurance	QA enforcement
Anchor	content-specialist	Content creation
Muse	creativity	Creative tasks
Cipher	knowledge-curator	Knowledge management
Lens	multimodal	Multimodal processing

Memory Architecture

Three system-level memory types:

• Episodic -- Failures and RCA reports
• Semantic -- SHARED_KNOWLEDGE.json
• Procedural -- PROCESSES.json

Vector embeddings stored in ~/.openclaw/memory/lancedb/.

Full guide: See openclaw-memory skill for the four-layer memory model (bootstrap files, daily logs, session transcripts, retrieval index), shared memory access patterns, SCRIBE compression, and the /oc-memory command.

Key Principles

• Never hardcode absolute paths
• Keep agent data within agent workspaces
• Place shared data in workspace root
• Secure sensitive files with appropriate permissions (chmod 600 for auth)
• Maintain self-contained skills with SKILL.md definitions
• Server parity via /home/openclaw/ using Tailscale

When Auditing

To inspect the workspace:

# Show top-level structure
ls -la ~/.openclaw/

# List agent workspaces
ls ~/.openclaw/workspace/agents-workspaces/

# Check a specific agent's files
ls -la ~/.openclaw/workspace/agents-workspaces/<agent-id>/

# Verify auth file permissions
stat ~/.openclaw/workspace/agents-workspaces/*/auth-profiles.json

# Check shared knowledge
cat ~/.openclaw/workspace/SHARED_KNOWLEDGE.json | head -50

# Inspect communication channels
find ~/.openclaw/workspace/comms/ -type f