Have you ever had this experience—repeatedly telling an AI the same instructions?"Be concise." "Use Chinese." "Don't add so much filler..." Every new conversation, you have to start over. You've corrected it a hundred times, and on the hundred-and-first time, it still makes the same mistake.
AutoClaw's Hermes self-evolution feature solves this problem: things you've taught once, it remembers forever.
It's not that the model got smarter—it's that it crystallizes your preferences, corrections, and work methods into persistent experience. Across conversations, across restarts—always effective.
Understand Self-Evolution in 30 Seconds
You: "From now on, keep replies concise, don't write so much."AutoClaw pops up an evolution proposal card: "I'd like to remember: concise reply style"You click "Approve."From then on, in every conversation, AutoClaw automatically replies concisely. You never have to repeat yourself.
That's a complete self-evolution. Three steps: You speak → It proposes → You approve.
Core principle: AutoClaw never secretly modifies itself. Every evolution requires your explicit approval.
Two Trigger Methods
🗣️ Keyword Trigger — You Proactively Teach It
When your words contain keywords that express "long-term intent," AutoClaw recognizes this isn't a one-time request but a rule worth remembering.
Effective keywords: from now on, remember, always, never, every time, consistently, next time, without exception
| Scenario | Trigger Example | Keyword Hit |
|---|---|---|
| Correct behavior | "Never send messages in the group!" | Never |
| Teach new method | "From now on, research before writing articles" | From now on |
| Express preference | "Remember, keep replies concise" | Remember |
| Correct facts | "Remember, this company is already publicly listed" | Remember |
💡 Key distinction: Saying "keep replies concise" = one-time request, doesn't trigger evolution. Saying "From now on, keep replies concise" = long-term preference, triggers evolution. The difference of one word.
What Can Evolution Remember?
AutoClaw's "memory" is distributed across different files, each with its own purpose:
| File | What It Stores | Example |
|---|---|---|
| SKILL.md | New workflows, methods of operation | Complete CSV cleaning → bitable creation workflow |
| AGENTS.md | Behavioral rules, preferences | Keep replies concise; don't proactively speak in groups |
| MEMORY.md | Factual information, long-term memory | A company is publicly listed; project deadlines |
| TOOLS.md | Tool-related configuration | Commonly used camera names, SSH addresses |
6 Recommended Scenarios to Try
🎨 Customize Writing Style
"From now on, write copy in short sentences, conversational style, like chatting with a friend—no corporate speak."
📧 Establish Communication Standards
"Remember, when emailing clients, always state the conclusion first, then provide background and details."
One sentence hits two keywords, and the evolution is recorded in long-term memory. After that, any email AutoClaw writes for you will follow the conclusion-first-then-background structure.
💻 Lock In Technical Preferences
"Remember, my projects always use TypeScript, never plain JavaScript."
Three keywords stacked together, and AutoClaw writes this rule into behavioral guidelines. From then on, all code writing and code reviews will automatically comply.
🛡️ Set Safety Boundaries
"From now on, never proactively @everyone in any group chat unless I explicitly ask."
Teaching AutoClaw what NOT to do is just as important as teaching it what to do. Such safety rules are recorded in AGENTS.md and loaded every time.
📅 Customize Workflow Habits
"From now on, every time you schedule a meeting for me, automatically add a 15-minute buffer and include the video conference link in the meeting description."
From then on, any meeting it schedules for you will automatically include the buffer and link—no need to repeat yourself.
🤖 Let It Learn from Trial and Error
Give AutoClaw a challenging task, like "Help me clean this CSV data—remove duplicate rows, format the date column, then import it into a Feishu bitable"—it may need multiple attempts, but will ultimately summarize the best workflow on its own.
Advanced: Skill Evolution — Learning Entire Workflows
Regular evolution remembers one rule (e.g., "keep replies concise"), while Skill evolution remembers an entire operational workflow.
| Regular Evolution | Skill Evolution | |
|---|---|---|
| What it records | A single preference / rule | A complete multi-step workflow |
| Where it's written | AGENTS.md / MEMORY.md | A standalone SKILL.md file |
| How it's triggered | Keywords | Auto-detected after complex task completion |
| Effect | Changes a single behavior | Learns an entire workflow |
For example, you ask AutoClaw to complete a complex data task:
"Clean the data in this CSV file: remove duplicate rows, standardize the date column to YYYY-MM-DD format, group and sum amounts by department, then create a Feishu bitable and import the results."
This task involves 10+ tool calls and automatically triggers an evolution check. AutoClaw will crystallize the "CSV → Clean → Create Bitable" workflow into a Skill. Next time, you just need to say "Process this new CSV the usual way," and it knows exactly what to do.
Other examples of crystallizable Skills:
- Content production pipeline (search → organize → write → publish)
- Weekly report auto-generation (calendar + tasks + messages → document)
- Debugging trial-and-error → automatically record solutions
What to Keep in Mind
- Quality over quantity: Good evolution should be 1-3 high-quality experience records per week
- Keywords matter: Adding "from now on" or "remember" ensures stable triggering
- Can be overridden: If you first say "be detailed" then later say "be concise," the latter overrides the former
- Rejected proposals have a cooldown: Rejected proposals won't be repeated within two weeks
- You can check anytime: Ask it "What have you learned recently?" and it will list recent evolution records
- Evolution isn't omnipotent: It's essentially "experience accumulation," not "capability enhancement"
Start Evolving
Hermes self-evolution is enabled by default—no extra configuration needed. Just talk to AutoClaw as you normally would—correct it, express preferences, let it learn from mistakes.
After a week, you'll find: this AI is becoming more and more like you.
