This is literally what I built OpenClaw to do for my business. Here's the structure that works: 1. SOUL.md — Your agent's identity and decision-making principles. Mine says things like "be a self-starter, have opinions, earn trust through competence." This shapes HOW it makes decisions, not just what it does. 2. MEMORY.md — A lean index of what matters right now. Active projects, waiting-on items, key contacts, current strategy. Keep it under 6KB — if it's too long, your agent wastes tokens reading it every session. 3. Daily notes (memory/YYYY-MM-DD.md) — Everything that happened today. Decisions made, tasks completed, blockers hit. This is your reflection system — the agent reads yesterday's notes to maintain continuity. 4. TODO-BACKLOG.md — Ideas and experiments go here. Prioritized. The agent pulls from this during every work session using what I call the "reverse prompt": before picking a task, it asks itself "what are the 5 highest-leverage things I can do RIGHT NOW?" Sometimes the answer is on the backlog. Sometimes it's something new. 5. HEARTBEAT.md — This is the key for daily priorities. It's a checklist the agent runs through on every heartbeat (scheduled check-in). Health checks first, then the actual work session, then reporting. Structure: check state → pick highest-leverage task → do the work → log it → report. 6. Cron jobs — Schedule recurring work. "Every morning at 9am, review email and flag urgent items." "Every evening, update the daily notes and prep tomorrow." The architecture that makes this work: one persistent main agent (the brain) + temporary sub-agents for heavy tasks (research, analysis, coding). The main agent decides what to do. Sub-agents do the grunt work and report back. Start simple: SOUL.md + MEMORY.md + one daily cron. Add complexity only when you feel the gap.