The System That Worked (For 3 Days)
Like every builder with ADHD, I've tried every productivity system. GTD. Bullet journals. Notion databases. Obsidian vaults. They all share the same fatal flaw: they require consistent manual maintenance.
My latest attempt was a Chief of Staff (COS) system inspired by Nate's Delegation Kit. Four markdown files synced across devices:
PROJECTS.md- Active initiatives with status and next actionsINBOX.md- Unprocessed captures waiting for triageWAITING_FOR.md- Things other people owe meDECISIONS.md- Choices made and why
Day 1: Everything was current. Zero overdue items. Clean system, clean mind.
Day 3: INBOX had 3 unprocessed items. Not bad. I'll get to it.
Day 7: PROJECTS.md diverged from reality. 12 items overdue in WAITING_FOR.
Day 14: INBOX had 30+ items from 3 weeks ago. I stopped opening it.
Day 19: 50+ overdue items. PROJECTS.md referenced deals that had been dropped 2 weeks prior. The system was fiction. I was making decisions from memory, not from the system designed to replace memory.
Manual maintenance requires consistent discipline. ADHD brains don't do consistent. The person who needs the system most is the person least likely to maintain it.
The Insight That Changed Everything
I was already capturing everything to Open Brain (a semantic memory layer). Session summaries. Meeting notes. Person profiles. Commitments. Decisions. All of it was going into OB automatically through auto-capture skills at the end of every AI session.
The data was there. It just wasn't being turned into anything useful.
Then I read how Manus handles state. Their agents don't maintain a todo list. They regenerate it every cycle. The file is rewritten from scratch each time, pushing the current state into the model's attention window. The file is disposable. The data source is persistent.
"The agent keeps and rewrites a todo.md file throughout the task, checking off completed items and restating what remains. This constantly pushes the global plan into the most recent part of the context, where attention is strongest." - Nate's analysis of Manus
That's when it clicked: COS files are views, not state.
The Architecture
State lives in Open Brain. It's persistent, semantic, searchable. COS files are generated from that state. They're disposable. They get regenerated every 30 minutes. Nobody maintains them.
The flow is a four-node flywheel:
- Capture - Session summaries, meeting notes, voice transcripts all flow into Open Brain automatically. No manual step.
- Store - Open Brain holds 221+ thoughts with semantic search. Persistent across all sessions, all devices, all AI tools.
- Query + Generate - A life-engine cron runs every 30 minutes. It queries OB with semantic searches ("commitments waiting on others", "project status updates", "decisions made"). An LLM processes the results into a structured JSON state file, then generates the COS markdown files.
- Steer Attention - The fresh COS files are loaded into the next AI session via hooks. The model sees current reality, not 19-day-old fiction.
Three Design Principles (From Nate)
Don't build a complex schema. Models improve quarterly. Use semantic search, not SQL. The model's understanding of language IS the query engine. "Sarah owes me the proposal" gets correctly classified as a WAITING_FOR item without a due_date field.
COS files on disk are the right output format. Agents read files. Humans read files. Don't replace files with a dashboard or database query. Files are the universal interface.
The canonical state is a JSON file (cos-state.json). Models are less likely to corrupt structured JSON than freeform text. The markdown COS files are generated FROM the JSON. Machine-readable state, human-readable views.
The Bidirectional Trick
One concern: what if someone wants to edit a COS file directly? Add a quick WAITING_FOR item by hand?
The life-engine handles this. Before regenerating, it diffs the current COS files against cos-state.json. If it finds items in the markdown that aren't in the JSON, someone edited manually. Those new items get captured to Open Brain first, then the regeneration includes them.
Edits flow both directions. Open Brain is always the source of truth, but humans can still write to the files. Nothing is lost.
What Gets Killed
INBOX.md is gone. Auto-capture already does this better. Every session summary, every ACT NOW item from panning-for-gold, every commitment detected in a meeting transcript gets captured to Open Brain without manual triage. The inbox was a processing queue for a human who wasn't processing. Remove the queue, keep the capture.
What This Means for Service Businesses
If you're running an MSP, law firm, or any service business with AI tools, you probably have the same problem: the AI produces great analysis, but the operational files it references go stale because nobody updates them.
The pattern is universal:
- Make your AI memory layer (Open Brain, or any semantic store) the source of truth
- Generate your operational files from memory, don't maintain them manually
- Run the generation on a cron (30 minutes works for most workflows)
- Let humans edit the generated files, but sync edits back to the source
The system that works for ADHD builders is the system that doesn't require them to do the thing they can't do: maintain consistency over time. Let the AI do the maintaining. You do the thinking.
Build It Yourself
The full architecture, research, and overnight agent prompt are open:
- OB1 (Open Brain) - The semantic memory layer
- The COS-from-OB recipe will be submitted as an OB1 community contribution
- Total build time: ~2.5 hours from zero to working system
COS files are views, not state. Open Brain captures. Life-engine regenerates. You never maintain.