Why Saying "Hi" to Your Bot Eats 12% of Your Daily Quota (And How to Fix It)

A community member recently posted in Discord:

"I just sent a 'hi' to my claw bot and that consumed 12% of my daily usage wtf 😓 any tips on how I can optimise this?"

This catches many new users off guard. You send a simple greeting, expecting minimal usage, and suddenly a chunk of your daily quota is gone. Here's what's happening and how to optimize.

Why Simple Messages Are Expensive

When you send a message to your OpenClaw agent, it's not just sending your message to the model. Every request includes:

• System prompt - Your agent's core instructions (AGENTS.md, SOUL.md, etc.)
• Memory files - MEMORY.md, daily memory files, any loaded context
• Skill definitions - All available skills and their descriptions
• Recent conversation history - Previous messages in the session
• Tool schemas - Definitions for all enabled tools

All of this gets sent with every single message. That simple "hi" is riding alongside potentially 50-100K+ tokens of context.

How to Reduce Per-Message Cost

1. Trim Your System Prompts

Review your workspace files:

• SOUL.md - Does it need to be that long? Keep personality concise.
• MEMORY.md - Archive old info to dated files, keep active memory lean.
• AGENTS.md - Remove instructions the agent has internalized.

2. Use a Smaller Model for Simple Chats

Not every message needs Opus 4.6. Consider:

• Switch to Sonnet for casual conversation
• Use Flash models for quick lookups
• Reserve expensive models for complex tasks

You can switch models mid-session with /model in most channels.

3. Disable Unused Skills

Every skill in your skills/ directory adds to the context. If you're not using 10 of them, they're still being loaded. Move unused skills out of the active directory.

4. Review Loaded Context

Check what's being injected:

openclaw status --verbose

Look for unexpected files being loaded. Sometimes old configs or large reference docs get pulled in.

5. Start Fresh Sessions

Long-running sessions accumulate history. Each new message carries more weight as the conversation grows. Start fresh sessions for unrelated tasks.

The Tradeoff

Remember: all that context is what makes your agent useful. It knows who you are, remembers your preferences, and has access to tools because of that context. The goal isn't zero context—it's appropriate context.

A quick "hi" probably doesn't need your full memory file. But when you ask "what was that restaurant we discussed last week?", you'll be glad it's there.

Quick Win: Lighter Default Model

The fastest fix: set a lighter model as your default for general chat, and explicitly invoke heavier models when needed:

openclaw configure
# Set defaultModel to something efficient like sonnet-4.6 or gemini-3-flash

Then when you need heavy reasoning, switch with /model opus for that specific task.

Tip from the OpenClaw Discord community. Join us to share your own optimization tricks!