How to set up OpenClaw Agents that actually get better Over Time (My exact stack after 40 Days)

三德子曰：此文所述，以恒久对话养智能，不追逐速变，此乃慢钱胜快钱之真谛也。其化繁为简，以牍记（markdown files）累积智慧，如日积月累之账簿，贵在持之以恒，而非朝夕之功。然，人机对话亦耗心力，反馈之精妙，或存偏颇。吾辈当思，系统虽简，其维系仍需细致，方能避短期之躁进，求长久之稳健，保生存之道，胜一时之利。

My agents get smarter every day. All I do is talk to them.

Not tweak prompts. Not swap models. Not rebuild the architecture.

Just talk. Give feedback. Watch them write it down.

40 days ago, my content agent drafted tweets with emojis and hashtags. My research agent buried signal in noise. I was spending more time correcting them than the tasks would have taken me to do myself.

Today Kelly drafts in my exact voice. Dwight delivers 7 stories every morning, every one worth reading. Eight agents running 24/7. I open Telegram, review drafts, drink my coffee.

Same model on day 1 and day 40. The difference is a stack of markdown files that get richer every single week.

This is that stack.

The stack

Three layers make up the entire operating system:

That's it. No orchestration framework. No message queues. No database. Markdown files on disk. The filesystem is the integration layer.

Layer 1: Identity

SOUL.md (who the agent is)

It defines who the agent is, what it does, and how it behaves.

Here's a trimmed version of Dwight's, my research agent:

The TV character trick. Every agent is named after a TV character. When I tell Claude "you have Dwight Schrute energy," it already knows what that means from training data. Thorough, intense, takes the job dead seriously. That's 30 seasons of character development loaded for free.

Keep it under 60 lines. SOUL.md loads every session. If it's too long, it eats context that should go to actual work. Identity, role, principles, relationships, vibe. That's all you need.

Here's a starter template:

Start with one agent. Pick your most repetitive daily task. Write a rough sketch. The first version will be mediocre. You'll rewrite it ten times over the next month based on what you see.

IDENTITY.md (the quick reference card)

SOUL.md is the full personality. IDENTITY.md is the business card. Name, role, vibe, one-liner.

Small file. Big quality-of-life improvement when you're running 8 agents. This is what shows up in Telegram when agents message you.

USER.md (who the agent works for)

Every agent needs to know who it's helping. USER.md holds your preferences, your background, and the context that shapes how the agent behaves.

Write it once. Every agent reads it.

The personal details matter more than you'd think. Timezone means agents don't schedule things at 3 AM. Dietary preferences mean when Pam drafts a newsletter about a team dinner, she doesn't suggest a steakhouse. These details compound.

Layer 2: Operations

AGENTS.md (Behavior rules)

SOUL.md is who the agent is. AGENTS.md is how it operates. Session startup routines, file reading order, memory management, safety rules.

Here's the root-level AGENTS.md that every agent inherits:

Then each agent adds its own. Kelly's AGENTS.md extends this with her specific workflow:

Agents have no memory between sessions. Everything starts fresh. If a correction doesn't reach a file, it doesn't exist next session. AGENTS.md makes this explicit so the agent writes everything down.

Specialist files are where agents get sharp. Kelly doesn't just have AGENTS.md. She has six extra files that define exactly how she creates content: writing style guides, post format references, real examples, daily assignments.

Dwight has a target audience profile and a research protocol. Each agent's folder grows as the role gets more defined. Start with AGENTS.md.

Add specialist files only when you notice a pattern that keeps needing correction.

HEARTBEAT.md (for self-healing)

Agent teams are infrastructure. Infrastructure breaks.

Monica's HEARTBEAT.md:

Monica runs this on every heartbeat. She checks two things: whether the browser is alive, and whether the cron jobs actually ran.

They're connected. If the browser dies, Dwight can't do his research sweeps. If Dwight misses a sweep, Kelly and Rachel draft content from stale intel. If the cron jobs silently stop running, the whole operation looks healthy on the surface while nothing is actually happening.

That last one is exactly what hit me in week three. The scheduler had a bug. Jobs were advancing in the queue but never executing. I didn't notice for hours.

After that, I built the heartbeat to catch both failure modes in one place. It has, multiple times since.

You don't need this on day one. Build it after your first failure. You'll know exactly what to monitor because you'll have felt what breaks.

Layer 3: Knowledge

The memory system that works is a three-tier system built on files.

Tier 1: MEMORY.md (curated long-term memory)

Not raw logs. Not everything that ever happened. The stuff that matters.

From Monica's MEMORY.md:

Notice the "Hard Lessons" section. Monica deleted a project folder. Now that mistake lives in her long-term memory permanently. She'll never do it again. One correction, stored once, preventing the same error across every future session.

From Kelly's MEMORY.md:

Kelly wrote the "BAD" section herself after corrections. She catalogues her own mistakes so she doesn't repeat them. That section alone is worth more than any prompt engineering guide.

Security note. MEMORY.md only loads in direct sessions, not shared contexts like group chats. Keep sensitive preferences out of files that load everywhere.

Don't write MEMORY.md on day one. It grows from feedback. Give feedback → agent logs it in daily memory → distill the important stuff into MEMORY.md → it loads every session → the correction never needs to be given again.

Tier 2: memory/YYYY-MM-DD.md (daily session logs)

Raw notes. What happened today. What was drafted. What feedback came in.

Daily logs are the raw material. MEMORY.md is the refined product. You need both.

The maintenance rule. Daily logs accumulate fast. If you don't prune them, your agent's context balloons. Kelly's hit 161,000 tokens. Output quality tanked. I had to compact her to 40,000. Now I review and archive old daily logs every two weeks.

Only load today's log plus yesterday's. The agent doesn't need its entire history every session.

Tier 3: Organized memory folders

At the root level, I organize memory by person:

Organize by person or project as your setup grows.

Shared Context (cross-agent knowledge layer)

This is the newest addition and the one that changed everything. A single folder that every agent reads at session start.

THESIS.md is my current worldview. What I care about, what I've already written, and what gaps remain. Dwight reads it to prioritize research. Kelly reads it to match my thinking. Ryan reads it to propose articles. Every agent aligns to the same source of truth.

FEEDBACK-LOG.md is the cross-agent correction layer. When I tell Kelly "no em dashes," that feedback applies to Rachel, Ryan, and Pam too. Instead of correcting four agents individually, I write it once and every agent reads it.

How agents coordinate

No API calls between agents. No message queues. Just files.

Dwight writes research to intel/DAILY-INTEL.md. Kelly reads it. Rachel reads it. Pam reads it. The coordination is the filesystem.

One agent writes. Other agents read. The handoff is a markdown file on disk.

The one-writer rule. Never have two agents writing to the same file. Design every shared file with one writer and many readers. This prevents every coordination conflict you'd otherwise have to debug.

Scheduling makes this work. Dwight runs at 8 AM and 4 PM. Kelly and Rachel run at 5 PM. Dwight runs first because everyone depends on his output. Get the order wrong and downstream agents read stale or empty files.

The full directory structure

Why this works

The files aren't static. They evolve.

Kelly's SOUL.md on day one was a rough sketch. By day 40, it has specific voice examples, a list of rejected patterns she wrote herself, and a "NEVER SUGGEST AGAIN" section with every topic she's already covered.

Dwight's principles on day one said "find what's trending." By day 10, they said "If Alex can't DO something with it TODAY, skip it." (Alex is our target reader profile, the developer we build content for.) By day 20, he'd added verification steps: check repo creation dates, check Show HN timestamps, trace X discoveries to primary sources.

The shared-context layer didn't exist until day 20. I was repeating the same corrections to multiple agents. Then I built THESIS.md and FEEDBACK-LOG.md, and suddenly one correction propagated everywhere. That single change saved me more time than any prompt optimization ever had.

The model is the same on day 1 and day 40. It doesn't get smarter because you've been using it longer. But the files around it get richer, sharper, more specific to your exact needs. That accumulated context is the moat. Nobody can replicate it by using the same model.

You earn it by showing up and talking to your agents every day.

How to start

Don't build all of this in a weekend. I didn't.

Today. Install OpenClaw. Write one SOUL.md, one IDENTITY.md, one USER.md. Pick your most repetitive daily task. Set up one cron job. Let it run.

After 3 days. Your agent's output will be mediocre. Start giving specific feedback. Make sure the feedback lands in a memory file, not just the chat.

After 1 week. Create AGENTS.md. Define the session startup routine. Add the memory management rules.

After 2 weeks. Start MEMORY.md. Review your daily logs. Which corrections keep recurring? Distill them into permanent entries. This is when you'll feel the compounding start.

After 3 weeks. Add your second agent. Set up file-based coordination: first agent writes to a shared file, second agent reads it. Add role-specific guides as patterns emerge.

Around the same time. Build the shared-context layer. You'll feel the pull before you get here. Repeating the same correction to multiple agents is the signal. THESIS.md for your current thinking. FEEDBACK-LOG.md for cross-agent corrections.

After 4 weeks. Add HEARTBEAT.md after your first failure. You'll know exactly what to monitor because you'll have felt what breaks.

All you have to do is talk to your agents. The files do the rest.

If you haven't read the first article yet, I'd highly recommend to do it now.

I'll be publishing more about OpenClaw, autonomous AI agent teams, and the evolving landscape for AI PMs and developers.

Follow me @Saboo_Shubham_ to stay tuned.

本文由 sandezi-page 自动发布