templates

Reference Config Starter Kit

A working-by-default starter bundle that matches the patterns documented in the OpenClaw Optimization Guide. Copy these files into a fresh OpenClaw workspace and you're running the guide's baseline setup.

Tested on OpenClaw 2026.4.15 stable (2026-04-16). See Part 26 — Migration Guide if you're on an older version.

What's in this directory

File	What it is	Target size	Introduced in
openclaw.example.json	Reference config. Copy to ~/.openclaw/openclaw.json (or your project root), replace ${...} env var refs with real credentials, edit to taste.	~2 KB	Part 15, Part 24
SOUL.md	Agent personality file (tone + anti-patterns). Injected every message.	< 1 KB	Part 2
AGENTS.md	Operating rules: decision tree, orchestration, approval categories, safety. Injected every message.	< 2 KB	Part 5
MEMORY.md	Pure index with links into vault/. Injected every message.	< 3 KB	Part 4
TOOLS.md	Tool name + one-liner per line. Stops model re-reading full tool schemas.	< 1 KB	Part 2
vault/projects/	Project-level notes directory (one folder per active project).	—	Part 9

Combined injected-context footprint target: < 8 KB. Compare against benchmarks/.

30-second install

# 1. Back up your existing config
cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.pre-guide.$(date +%Y%m%d) 2>/dev/null || true

# 2. Drop the reference config in (after replacing ${...} with your real keys)
cp templates/openclaw.example.json ~/.openclaw/openclaw.json

# 3. Drop the injected-context files into your workspace
cp templates/SOUL.md templates/AGENTS.md templates/MEMORY.md templates/TOOLS.md ./

# 4. Restart the gateway
openclaw gateway restart

# 5. Verify
openclaw doctor

Full setup in setup.sh / setup.ps1 at the repo root.

Philosophy (one screen)

1. Target < 8 KB of injected context. Every KB above that pays tax on every turn. See Part 1 — Speed and Part 2 — Context Engineering.
2. MEMORY.md is an index, not a store. Details go in vault/. Memory-core's built-in dreaming keeps it small automatically. See Part 4 + Part 22.
3. Orchestrator + workers, not a megamind. Frontier model plans, cheap workers execute. See Part 5 + Part 24.
4. Semantic approvals, not name allowlists. Read-only allow, execution ask, write.network ask, control-plane deny. See Part 24.
5. Never commit credentials. Use env var refs in openclaw.json. See Part 15.

When this kit does not match what you should run

• Single-agent personal-dev setup: you can skip the Task Brain approval block and pin one model.
• Local-only setup on a 14B-or-smaller model: flip agents.defaults.experimental.localModelLean: true in openclaw.json.
• Team deployment with multi-user approvals: override taskBrain.approvals per-agent — see Part 24.
• Large vault (500+ files): add LightRAG (Part 18) and Repowise (Part 19).