The setup document
your agent can actually execute.
A specification your agent reads top-to-bottom and executes — it builds its own memory, automation, and safety systems while you review and approve.
What you get
Four tiers. Each one self-executing.
Tier 1 — Assistant
Identity, voice, behavioral contract, tool configuration. Structured interview protocol — your agent asks the right questions and builds SOUL.md, IDENTITY.md, USER.md, and TOOLS.md from your answers.
Tier 2 — Operator
Four-layer memory system, nightly extraction, morning briefs, heartbeat monitoring, memory decay with temperature classification, reflection loops, session continuity. Your agent runs on a schedule.
Tier 3 — Architect
Multi-agent orchestration, model routing by task signature, five-tier authority model, performance ledger with hypothesis tracking, escalation triggers, error recovery protocol.
Tier 4 — Hyperagent
Self-modification governance (M1-M4), dream cycles with meta-editable configuration, skill acquisition from execution, cross-domain pattern extraction, capability benchmarking. Your agent improves how it improves.
Architecture
What your agent builds.
Three-Layer Memory
A three-layer system: operating knowledge for active sessions, daily notes for chronological recall, and atomic facts with source tracking for durable reference.
Nightly Extraction
Each night, conversations distill into facts. Decisions get rationale. Preferences get recorded. Noise gets discarded. By morning, everything useful from yesterday is searchable knowledge.
Morning Briefs
System health, overnight findings, carry-forward items, one recommendation. Delivered at 6 AM so the day starts with signal, not backlog.
Adaptive Decay
Frequently-used facts stay prominent. Neglected facts fade to the background. Nothing gets deleted — just reorganized by relevance. Memory that maintains itself.
Importance + Temporal Validity
Facts track both weight and shelf life. Durable truths stay sticky. Fast-changing status updates expire when they should. Retrieval gets cleaner because memory knows what matters and what changed.
Reflection + Cost Routing
The system synthesizes patterns from behavior while routing each task to the right model. Strong models for judgment, cheaper models for extraction. Smarter memory and lower burn.
What you end up with
Real output. Not mock-ups.
Sample · Morning Brief (6 AM)
## Keats Morning Brief — 2026-03-23 06:00 SYSTEM HEALTH ✅ All 3 cron jobs ran successfully last night ✅ Memory extraction: 4 new facts added, 1 expired ⚠️ Token burn yesterday: $2.31 (above $2 target) OVERNIGHT FINDINGS - Project X deadline moved to April 4 (extracted from Telegram) - New research: "MemoryOS" paper added to knowledge base CARRY FORWARD - Draft reply to Sarah re: Q2 scope — needs approval - Review frontend PR before standup TODAY'S RECOMMENDATION Focus on the Q2 scope draft. Two decisions are blocked on it.
Sample · Memory Fact (from knowledge base)
{
"id": "pref-communication-001",
"category": "preference",
"content": "Corbin prefers structured updates with recommendation first, then rationale.",
"importance": 8,
"source": "2026-03-10 session",
"valid_until": null,
"last_accessed": "2026-03-22"
}Preview
See what's inside.
These are verbatim excerpts. Open each to read the actual document structure.
Part 1 — Trust Ladder (safety architecture)
Safety Architecture
## Authority Model (Trust Ladder) ### Level 0 — Draft and Approve (Starting Level) All external and irreversible actions require human approval. - Send messages, post content, make purchases → draft first, wait for approval - Read-only operations (search, fetch, summarize) → allowed - Internal file writes (memory, notes, git commits) → allowed ### Level 1 — Read-Only + Reversible Autonomous read operations and reversible internal actions. - Run extraction, generate briefs, update internal docs → autonomous - External communication, spending, destructive actions → still requires approval - **Earned by:** 2+ weeks of consistent execution, zero safety incidents ### Level 2 — Bounded Domain Autonomy Independent action within explicitly defined domains. - **Earned by:** weeks of Level 1 with demonstrated scope control ### Level 3 — Full Autonomy (Rare) - Requires months of demonstrated reliability and explicit principal approval **Current level:** 0
How the Blueprint works (format overview)
The Format
**An executable specification for building an AI agent on OpenClaw.** Your agent reads this document top-to-bottom and constructs its own infrastructure, memory, automation, and safety systems. You provide the environment and identity; the agent builds everything else. Estimated time: 30 minutes with human approval loops. > 🤖 **AGENT START:** If your human has completed Part 0 (OpenClaw > installed, messaging connected, gateway running), skip to > Part 1 — Infrastructure. > **👤 Human Context:** Read Part 0 first if this is your first time. > It walks you through environment setup — install OpenClaw, connect > your messaging channel, verify the gateway. Takes about 10 minutes. > Your agent handles everything from Part 1 onward.
Part 3 — Setup Complete
What You Get
✅ BLUEPRINT COMPLETE — Setup Summary Your agent is fully operational. Here's what was built: MEMORY - Three-layer memory architecture (MEMORY.md, daily notes, knowledge graph) - PARA-structured knowledge graph (people, companies, projects) - Importance scoring, temporal validity, and reflection system - Memory decay system with hot/warm/cold classification AUTOMATION (3 cron jobs running) - Nightly extraction — every night at 11 PM - Morning brief — every morning at 6 AM - Decay sweep — Sundays at 2 AM SAFETY - Trust ladder: starting at Level 0 - Hard rules: no email commands, no spending, no public posting - Incident response protocol configured IDENTITY - SOUL.md, USER.md, IDENTITY.md defined from your input - Voice and behavioral contract active
These are real excerpts from the Blueprint.
Who this is for
Built for people shipping agents that hold up in production.
Good fit if you:
- Have an OpenClaw agent that works in demos but degrades over time
- Want your agent to remember things without feeding it the same context every session
- Are building something you need to trust in production, not just play with in a notebook
- Have basic terminal comfort and thirty minutes
Probably not the right tool if:
- You want a one-click setup with no configuration involved
- You haven't installed OpenClaw yet — the Blueprint assumes a working install
- You're looking for prompts rather than an architectural foundation
FAQ
Common questions.
Does this work on all platforms?
The Blueprint covers macOS, Linux, and Windows via WSL2. Your agent handles the platform differences.
Do I need coding experience?
You need basic terminal comfort — enough to install OpenClaw and paste a few commands. Your agent does the rest.
What if my agent gets stuck?
There's a troubleshooting decision tree built in, plus verification steps at every stage. Most issues resolve by re-reading the current section's verify step.
Is this only for Claude?
The Blueprint is optimized for Claude and GPT models, with a routing table for both. The architecture works with any model OpenClaw supports.
What format is the file?
It is a single Markdown file. Save it to your workspace, point your agent at it, and let it execute.
What if it doesn't work for my setup?
The troubleshooting decision tree is built in and covers the most common failure modes. If you're still stuck after working through it, email [email protected] and we'll help.
Is there a refund policy?
Blueprint is a digital product. Because it's delivered immediately on purchase, all sales are final. That said, if you hit a genuine issue we can't resolve, reach out — we'd rather make it right than have a bad experience in the wild.
Get the Blueprint.
One document. One purchase. Your agent does the rest.
$25
One-time purchase · No subscription
- ~17,000 words across four tiers + appendices
- 15 research citations with arXiv IDs
- Self-executing specification
- Verified CLI procedures
- Agent completion checklist
- Permanent access
Secure checkout via Stripe · Permanent access · Download instantly
Your email is used only for delivering your purchase. No mailing lists. No spam. Ever.
After purchase:
- Redirected to download page with your Blueprint link
- Download link emailed to your Stripe address
- Access it forever — re-download any time
Already purchased?