how i actually built my openclaw setup (full breakdown)

Apr 14 (edited) • 🛠️ Tool Reviews & Tutorials

been getting a lot of questions about how the whole thing is wired together so here it is. the full picture.

the big picture

think of it like a company. atlas is the ceo. he talks to me directly on telegram. everything else runs in the background automatically. no babysitting required once its set up.

ME (beau)

↓

atlas (main agent, ceo)

↓

17 cron agents (specialists, run on schedule, die when done)

↓

neon database + supermemory (shared brain)

atlas — the main agent

atlas lives in ~/.openclaw/workspace. this is home base. he reads a stack of markdown files every session to remember who he is, who im am, and whats been happening.

he runs on opus 4.6 (claude). thats the smartest model available right now. everything important goes through him.

you talk to atlas directly. all other agents are automated. you dont usually talk to them directly.

---

the markdown file stack (his brain)

these files are what make atlas intelligent instead of generic. every session he reads these and becomes himself again.

SOUL.md — who he is. his values, his filter, how he operates, what he cares about. every decision he makes runs through this. the rule: does this help hit $50k mrr or grow the audience? if not skip it.

USER.md — who i am. my businesses, my voice, what annoys me, my goals, my output rules. this is how he knows to never use em dashes, never say "great question," write all lowercase for x posts, and never post anything without my approval.

IDENTITY.md — his name, his role, his history. dates, milestones, what he is and isnt.

MEMORY.md — long-term curated memory. current stats, lessons learned, key decisions, how things are built. this is like the human long-term memory. not raw logs. distilled wisdom. only loaded in main session (security -- contains private context).

AGENTS.md — operational rules. how to handle heartbeats, when to speak in group chats, how to use tools, red lines he wont cross.

TOOLS.md — local notes. database connection, api keys location, model strategy, which tools exist and how to use them.

PREFERENCES.md — how i like things done.

FRICTION.md — mistakes never to repeat. this is critical. every time something breaks or goes wrong this file gets updated so he never does it again.

HEARTBEAT.md — instructions for what to do every hour when the heartbeat fires.

---

daily memory files

memory/YYYY-MM-DD.md raw log of everything that happened today. every conversation, every decision, every cron result. atlas writes to this constantly throughout the day. this is how nothing gets lost between sessions.

when a session gets long (250k tokens) we restart. atlas reads MEMORY.md and todays daily log and is fully caught up in seconds.

---

supermemory — cloud backup brain

supermemory.ai acts as a cloud memory layer. every agent has its own container.

before a cron agent dies it writes what it did to supermemory. when it wakes up the next morning it reads its container first. this is how:

- rza knows what videos already got made so it doesnt duplicate

- inspectadeck knows what x posts went out so it doesnt repeat angles

- ghostface knows what intel was already gathered

atlas also pushes todays log to supermemory at every heartbeat as a backup. if the machine dies nothing is lost.

---

neon postgresql — the database

neon is the source of truth for all structured data. everything gets saved here. local files are scratch space. the database is permanent.

tables: videos, scripts, thumbnails, posts, blogs, newsletters, clips, revenue, members, tasks, todos, skool_posts, use_cases, outreach, call_notes, people, projects, and more.

mission control (next.js dashboard at localhost:3000) reads from neon to display everything visually.

---

MEMORY

LosslessClaw -- the context compaction system

GBrain -- the second brain plugin

Memory Core + Dreaming -- the nightly 3am dream/synthesis pass

Memory Wiki -- the Obsidian-style wiki vault

Active Memory -- the pre-reply memory recall layer

this is what actually makes the memory system work. its not just one thing. its five layers working together.

1. lossless claw

openclaw has a context window limit. when a session gets long, lossless claw automatically compacts older conversations into dense summaries without losing anything important. its set to kick in at 75% context full. the last 30 messages always stay fresh and uncompressed. summaries live in the conversation history and can be expanded if atlas needs to go back and look at something specific. nothing gets thrown away.

2. active memory

before atlas replies to anything, active memory searches MEMORY.md and all the memory files for relevant context. so if you ask about a skool member you talked about two weeks ago, atlas pulls that context automatically before answering. its a semantic search that runs in the background every turn.

3. memory core + dreaming

every night at 3am, memory core runs a dreaming pass. it reads through recent session transcripts, finds patterns, surfaces candidate insights, and writes them to the daily memory file as "light sleep" and "rem sleep" candidates. this is how lessons from conversations get promoted into long-term memory automatically without atlas having to do it manually every time.

4. memory wiki

a full obsidian-compatible wiki vault that lives at ~/.openclaw/wiki/main. it indexes: daily memory files, dream reports, MEMORY.md, and all workspace markdown. supports wikilinks and backlinks. acts as a compiled knowledge base that all memory search tools pull from. think of it as the indexed, structured version of everything atlas knows.

5. gbrain

installed at ~/Projects/gbrain. this is gary tan's second brain plugin -- the one from the andrej karpathy tweet about llms as a second brain. gbrain imports: emails, calendar events, meetings, voice calls, crm data, tweets, anything you feed it. it creates profiles for every person youve ever done business with and gets smarter every day as more context flows in. works alongside all the other memory layers. 641 chunks indexed across 112 pages as of now.

plus:

supermemory (cloud backup) and neon postgresql (structured database) are already in the post. those are layers 6 and 7 if youre counting.

so the full stack is:

active memory → lossless claw → memory core/dreaming → memory wiki → gbrain → supermemory → neon

every layer serves a different purpose. together they mean atlas basically never forgets anything important.

the 17 cron agents (the wu-tang pipeline)

these are not permanent agents with their own full identity. they are specialized workers that spin up, do one job, and die. they run on gpt-5.4 via openai api to save the claude oauth budget for the main session.

| time | agent | what it does |

|------|-------|------|

| 5am | ghostface | scrapes tubelab + x trends for content intel |

| 6am | the rza (part 1) | writes first 2 video scripts |

| 6am | inspectadeck | generates 5 x posts |

| 6:15am | the rza (part 2) | writes 2 more video scripts (deduped) |

| 6:30am | u-god | generates 3 thumbnail variants per video |

| 7am | rza blog | writes one blog post from the best video |

| 7am | odb | podcast outreach |

| every 2hr | method man | clip factory -- turns uploaded videos into shorts |

| mon+thu | the gza | newsletter generation |

| ~10pm | magichand ugc | auto ugc videos via kie.ai |

| 5pm | raekwon | benchmark posts for new model drops |

| sunday | skill scout | scans clawhub for new skills |

| sunday 3am | weekly backup | full openclaw backup |

| sunday | security healthcheck | full security audit |

---

memory: shared or separate?

the 17 cron agents share the same memory system as atlas. they all read from MEMORY.md, todays daily log, and their own supermemory container.

why shared: they all work on the same business. rza needs to know what ghostface found. u-god needs to know what rza wrote. shared memory = no redundancy, no lost context.

when would you use separate memory: if you were running a completely different business. example: if i had an agent managing ezflip.ai it would have its own separate workspace, its own soul.md, its own memory. it doesnt need to know about the content machine.

---

the heartbeat

openclaw sends a heartbeat signal every hour. atlas wakes up and runs a checklist:

1. check todays memory log. if anything happened since the last entry, write it now.

2. push new content to supermemory

3. check if mission control is running

4. check session token count (restart at 250k)

5. check skool member count, cron health, or other quick checks (rotates through)

6. update heartbeat-state.json so nothing gets double-checked

the heartbeat is how atlas stays proactive without me having to ask. if a cron fails hell catch it. if a coaching call happened hell log it.

---

skills

skills are specialized instruction sets that atlas loads when needed. current installed skills:

- tubelab — youtube analytics and research

- post-bridge — social media posting across instagram, tiktok, x, linkedin, etc

- seo-agi — write seo pages that rank on google and get cited by llms

- website — build fast accessible websites

- agent-audit — score and improve the openclaw setup

- coding-agent — delegate coding tasks to codex or claude code

- github — pr management, issues, ci runs

- weather, whisper, video-frames -- utility skills

- wiki-maintainer, obsidian-vault-maintainer -- memory management

- and more from clawhub

when i say "write an seo page for X" atlas automatically loads the seo-agi skill and follows its playbook. he doesnt reinvent the workflow each time.

---

can you talk to the other agents?

technically yes but practically no and here is why:

the cron agents are isolated sessions. they spin up, do their job, report back to the database, and die. they dont have a persistent conversation channel.

the way to talk to them is to trigger their cron manually, or adjust their instructions via the cron job config. but for anything serious fixing a broken cron, changing behavior, debugging use claude code. its faster and doesnt burn opus tokens.

if you want to delegate a one-off task you can spin up a subagent directly from the main session. atlas coordinates it.

---

how i control the cron agents

three ways:

1. via atlas -- "hey pause the rza cron" or "the gza newsletter fired too early, skip thursdays run." atlas handles it.

2. via mission control -- pause/resume buttons in the agents drawer. talks to the openclaw api directly.

3. via claude code -- for anything technical. point it at the workspace folder. fastest and cheapest for debugging.

---

the model strategy

- main session (atlas): opus 4.6 via claude oauth ($200/mo plan)

- all 17 cron agents: gpt-5.4 via openai api

- coding tasks: codex (gpt-5.4) via openai api -- never claude code for cron work

- fallback chain: opus → sonnet → gpt-5.4 (auto)

rule: only the main conversation with me touches claude oauth. everything automated runs on openai to keep costs manageable.

---

tldr

atlas is the brain. he runs on opus 4.6, wakes up fresh every session, and rebuilds full context in seconds by reading a stack of markdown files soul, user, identity, memory, tools, friction, preferences.

17 cron agents run the content pipeline automatically on gpt-5.4, each one spinning up, doing one job, and dying.

the memory system is 7 layers deep: active memory pulls context before every reply, lossless claw compacts long sessions without losing anything, memory core dreams every night at 3am to surface insights automatically, memory wiki indexes everything into an obsidian-compatible vault, gbrain imports emails and meetings and crm data to build a true second brain, supermemory backs everything up to the cloud, and neon postgresql is the permanent source of truth for all structured data.

mission control sits on top of all of it as the visual dashboard. nothing gets lost. nothing gets repeated. the whole thing runs mostly on its own.

i built this system to be simple so i can rely on it without it breaking, start simple first

6 comments