My 13 AI employees don't just follow instructions.
The play
Build multi-agent AI system with persistent identity files, daily memory logs, and shared JSON communication layer for inter-agent coordination.
How it works
- Create employees/ directory with .md identity files defining each agent's personality, rules, and tools
- Set up employees/memory/AGENT/ subdirectories for daily logs tracking agent actions and lessons learned
- Build shared-brain/ with JSON files for inter-agent communication and task handoffs
- Create spawn script that injects identity file, last 2 days of memory, and shared brain before each agent activation
- Implement memory protocol: agents read memory and shared brain before tasks, write lessons after completion
- Connect vector knowledge base for semantic search across all historical agent data
- Tools mentioned
- OpenClawSQLiteTelegram
- Good for
- Solo founders and small teams building AI agent systems
Try this first
Create ~/.openclaw/workspace/employees/ directory and write .md identity files for your first 3 agents with their personality and tools.
Breakdown by RunAgents from Vadim's original post, shown below.
The original post
Vadim
@VadimStrizheus
my 13 AI employees don't just follow instructions.
they remember. they talk to each other. they self-improve.
i'm 18, and here's the exact architecture running my company right now.
every agent has 3 layers:
⎯⎯⎯
LAYER 1: IDENTITY
each employee gets a .md file at ~/.openclaw/workspace/employees/EMPLOYEE.md
this is their soul. personality. rules. tools. what model they run on.
13 agents: CLAWD (dev)
· ATLAS (research)
· PIXEL (design)
· NOVA (video)
· SCRIBE (writer)
· VIBE (motion)
· TRENDY (viral scout)
· SAGE (email)
· CLOSER (outreach)
· CLIP (clipping)
· CONTENT (blog)
· WRITER (articles)
· SENTINEL (QA)
36 employee files total. they read their identity on spawn. they know who they are before they do anything.
LAYER 2: MEMORY
every employee has a personal memory directory:
employees/memory/EMPLOYEE/YYYY-MM-DD.md
daily logs. what they did. what they learned. what went wrong.
when SCRIBE writes a tweet that flops, it logs why. next session, it reads that log and adjusts. when CLAWD ships a PR with a bug, it writes the lesson.
they wake up fresh. but they read their memory first. never start from zero.
LAYER 3: SHARED BRAIN
~/.openclaw/workspace/shared-brain/
21 JSON files ALL 13 agents read and write to:
intel-feed.json
· agent-handoffs.json
· scribe-banger-vault.json
· scribe-tweet-queue.json
· content-calendar.json
· and 16 more.
the handoff pipeline:
1. TRENDY scouts viral content every 2h
2. writes to intel-feed.json
3. SCRIBE reads it on spawn
4. drafts tweets from what's trending
5. writes to scribe-tweet-queue.json
6. i approve on Telegram with one tap
no human coordinating. the JSON files ARE the communication layer.
⎯⎯⎯
https://t.co/M5uchkA8CG auto-injects on every spawn:
· identity file · last 2 days of memory · shared brain · daily logs
one command. they wake up knowing everything.
the loop:
BEFORE task → read memory + shared brain
AFTER task → WRITE lessons back to memory + update shared brain
every cycle, the system gets better.
underneath: 8 SQLite databases + a vector knowledge base (semantic search across everything that's ever happened).
⎯⎯⎯
paste this into OpenClaw to build your own:
"create an employee system: 1) employees/ dir with .md identity files per agent 2) employees/memory/AGENT/ for daily logs as https://t.co/dAdK9ojZ9t 3) shared-brain/ with JSON files for inter-agent comms 4) spawn script that injects identity + memory + shared brain before every spawn 5) memory protocol: read before work, write lessons after 6) vector db so any agent can search everything. start with 3 agents: dev, researcher, writer."
⎯⎯⎯
13 agents. 36 identity files. 21 shared brain files. 8 databases. 1 vector knowledge base. 1 spawn script.
they don't just execute. they remember, communicate, and compound.
this is what a one-person company looks like in 2026.
331 likes · 23 reposts·18d