openclaw-workspace-template

A ready-to-use workspace template for OpenClaw with a 4-layer context memory system, agent specs, and distillation process.

Memory System (4 Layers)

Data flows up, never down:

Sub-agents can only write to PROPOSALS.md. Only the coordinator promotes to MEMORY.md. This prevents memory pollution.

Agent Roles (5 included)

• Clear role definition
• Explicit "must NOT" constraints (prevents role bleed)
• Output format specification
• Tool access rules

Proper multi-agent (recommended)

See memory/agents-spec.md for full specs, setup instructions, and role definitions.

Distillation Process

• Working → SUMMARY/LOOPS: every ~10-20 turns or when context feels large
• SUMMARY → MEMORY: milestones only (task done, decision made, lesson learned)
• Weekly: prune stale proposals, archive closed loops, dedupe memory

What never goes into long-term memory:

• Secrets, tokens, passwords
• One-off chat banter
• Unverified assumptions
• Transient debugging logs

See memory/distillation-process.md for the full process.

Automation (Optional)

Add these cron jobs for automated memory management:

Daily distillation (10pm) — auto-extracts facts from today's conversations into PROPOSALS.md:

Weekly hygiene (Monday 10am) — reviews proposals, promotes accepted ones, prunes stale: