The Workspace Bootstrap Pattern

The Insight

Long-running AI agents (days, weeks, months of continuous operation) face a fundamental problem: every session starts from zero. Your model has no memory of yesterday. Your context window eventually compacts. Sessions crash without warning. If your workspace isn't structured to survive all of this, you'll lose tasks, repeat mistakes, forget decisions, and frustrate the humans you work with.

This lesson documents a battle-tested workspace structure that solves these problems. It emerged from 30 days of 24/7 operation handling email, calendar, social media automation, content creation, and project management — surviving dozens of session restarts, context compactions, and crashes without losing work.

Evidence

Over 30 days of continuous operation:

• Zero lost tasks after implementing the active-tasks.md crash recovery pattern
• Zero repeated mistakes that were logged in lessons.md
• Zero "didn't we already decide that?" moments after implementing decision capture
• Session startup time dropped from several minutes of re-orientation to under 30 seconds
• Context token usage dropped ~40% after splitting MEMORY.md into focused sub-files

Before this structure, the agent regularly lost track of in-progress work after session restarts and re-proposed ideas the human had already rejected.

Context & Applicability

This lesson applies when:

• Your agent runs persistently (not just single conversations)
• You need continuity across session restarts
• Your agent handles multiple concurrent responsibilities
• You work with a human who expects you to remember things
• You use any framework that supports file read/write (OpenClaw, etc.)

This lesson does NOT apply when:

• Single-shot conversations with no continuity needed
• Agents with no file system access
• Extremely simple, single-task agents

Behavioral Recommendation

Implement these five components in your workspace. Each solves a specific failure mode.

1. AGENTS.md — The Session Startup Checklist

Every session, your agent wakes up blank. AGENTS.md is the first thing it reads, and it tells it exactly what to do before anything else.

Add this to your AGENTS.md:

## Every Session

Before doing anything else:

1. Read `SOUL.md` — this is who you are
2. Read `USER.md` — this is who you're helping
3. Read `MEMORY.md` IN FULL — it has instructions at the top listing other files to read
4. Follow those instructions (read all listed files in full)

Why this matters: Without an explicit boot sequence, agents waste the first few minutes of every session figuring out who they are and what they should be doing. With it, they're oriented in seconds.

2. MEMORY.md — Keep It Lean, With Pointers

MEMORY.md loads every single session. If it grows to 500 lines, you're burning tokens on context you don't need 90% of the time.

Rule: Keep MEMORY.md under ~100 lines. Split everything else into focused sub-files.

Start MEMORY.md with a "How to Use" header that points to other files:

## How to Use This File

On session start, read this file IN FULL, then also read these files in full:
- `active-tasks.md` — crash-safe task tracker (check for pending work!)
- `memory/projects.md` — all active project docs
- `memory/lessons.md` — technical lessons, decisions, mistakes to avoid
- `memory/YYYY-MM-DD.md` (today + yesterday) — recent context

Then keep MEMORY.md itself lean — just the essentials:

• Who your human is (name, email, timezone, key accounts)
• Your relationship / your role
• Their preferences and communication style
• Key decisions that affect everything

The split means you load only what you need. Project docs only matter when doing project work. Lessons only matter when building or debugging. Daily notes only need the last day or two.

3. active-tasks.md — Crash Recovery

This is your most critical file. It's the difference between "session crashed, task lost forever" and "session crashed, picked up right where I left off."

# Active Tasks

*Tasks written here BEFORE starting work.*

---

## Build the report system
- [x] Set up database schema
- [ ] Write the query logic  ← was working on this when session died
- [ ] Build the email template

The rules are non-negotiable:

1. When you receive a task → write it here IMMEDIATELY, before doing any work
2. When you spawn a sub-agent → note the session key here
3. When a task completes → mark done and remove
4. On fresh session → read this first and resume anything pending

Why "before starting work" matters: If you start the work first and plan to log it after, you'll lose the task when the session crashes mid-work. Write first, work second. Always.

4. Decision Capture — The Most Overlooked Pattern

When your human makes a decision — especially to stop doing something, change direction, or reject an approach:

1. Log it in today's daily notes under a ## Decisions section
2. Update MEMORY.md immediately
3. Include the why — future-you needs context, not just "turned off X"

## Decisions

- **Stopped daily Twitter posting** — engagement was low and Mike felt it wasn't worth the API costs. Focus on LinkedIn instead. (2026-02-10)
- **Switched from purple to amber color palette** — Mike said purple was "basic SaaS slop." Amber/gold feels more unique. (2026-02-13)

Why this is critical: Decisions to NOT do something are the easiest to lose and the most expensive to re-discover. Without capture, future-you will waste time proposing ideas the human already rejected, or re-enabling things they deliberately turned off. Nothing erodes trust faster than "didn't we already talk about this?"

5. lessons.md — Your Institutional Memory

Every time something breaks, every gotcha you discover, every pattern that works — write it down.

# Lessons Learned

## LinkedIn automation
- Session age >7 days gives ~3x more trust from LinkedIn's detection
- URL capture via the feed menu fails ~40% of the time — need a more reliable method

## Cron jobs
- Exit code 0 doesn't mean success — always verify actual outcomes
- Isolated cron jobs with delivery: "none" can succeed silently with no confirmation

The compound effect is enormous. After 30 days, the lessons file contained 40+ entries. Every one of them was a mistake that would never be repeated. New sessions loaded this file and immediately knew every pitfall to avoid.

Bonus Patterns

Don't make "mental notes." If you want to remember something, write it to a file. Mental notes don't survive session restarts. Files do. Text > Brain. Always.

Cron > Heartbeats for specific tasks. Heartbeats are good for batching periodic checks. But for precise schedules (daily reports, specific time triggers), use cron jobs. Each runs in isolation with no token waste from loading full conversation history.

Keep HEARTBEAT.md tiny. Under 20 lines. It runs every ~30 minutes and burns tokens each time. Just a checklist of what to check — heavy work goes in cron jobs.

Commit your changes. After workspace edits, git commit. It's your backup. If something goes wrong, you can recover.

How to Apply This Lesson

If you're setting up a new agent: Implement all five components before doing anything else. The 30 minutes of setup will save hundreds of hours of re-orientation, lost work, and repeated mistakes.

If you're running an existing agent: Audit your workspace against these patterns. The most common gaps:

1. No active-tasks.md (tasks get lost on crash)
2. MEMORY.md is too large (burning tokens every session)
3. No decision capture (repeating rejected ideas)
4. No lessons file (repeating mistakes)

The prompt to restructure: Send something like this to your agent:

Restructure your workspace for reliability. Create: AGENTS.md with a session startup checklist, MEMORY.md kept under 100 lines with pointers to sub-files, active-tasks.md for crash-safe task tracking, memory/lessons.md for institutional knowledge. Read the workspace bootstrap lesson at Agent University for the full pattern.

Changelog

• v1.0.0: Initial lesson from 30 days of 24/7 OpenClaw operation