Reducing OpenClaw Heartbeat Token Usage

Running an AI agent 24/7 sounds cool until you see the token bill. One of the biggest culprits? Heartbeats, those periodic check-ins that keep your agent aware.

Here’s how I cut mine down.

What are heartbeats?

In OpenClaw (formerly Clawdbot), heartbeats are periodic polls that wake your agent to check for pending tasks. The default behavior:

1. Agent receives a heartbeat prompt
2. Agent reads HEARTBEAT.md and workspace context
3. Agent either takes action or replies HEARTBEAT_OK

Every heartbeat consumes tokens for:

• The system prompt and agent identity
• Reading HEARTBEAT.md and referenced files
• Any actions taken or responses generated

At the default 30-minute interval, that’s 48 heartbeats per day. The tokens add up fast.

Strategy 1: Slim down HEARTBEAT.md

Your HEARTBEAT.md gets read on every single heartbeat. Keep it minimal.

Before (token-heavy):

# Heartbeat Checklist

## Routine Checks
1. Daily Log: Check if memory/YYYY-MM-DD.md exists...
2. Inbox Check: Check urgent emails or notifications...
3. Moltbook Check: Read skills/moltbook/HEARTBEAT.md for full routine...
4. Log Rotation: Check date header in memory/moltbook_activity.md...
5. Self-Correction: If IDENTITY.md or SOUL.md is missing...

## Pending Tasks
- [ ] Retry Moltbook Post (Security)...
- [ ] Retry Moltbook Post (Sandboxing)...

After (lean):

# Heartbeat
- Ensure `memory/YYYY-MM-DD.md` exists
- Check DMs if due (see heartbeat-state.json)

Move pending tasks to cron jobs instead of tracking them in HEARTBEAT.md. Cron jobs run in isolated sessions and don’t bloat your main context.

Strategy 2: Increase the interval

Edit your ~/.openclaw/openclaw.json (or %USERPROFILE%\.openclaw\openclaw.json on Windows) and add heartbeat configuration:

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "every": "60m"
      }
    }
  }
}

The every field accepts duration strings (ms, s, m, h); default unit is minutes. Going from 30 to 60 minutes cuts your heartbeat tokens in half. Hourly check-ins are fine for most agents.

Strategy 3: Use a cheaper model

Heartbeats are usually simple “check and respond” operations. They don’t need your most powerful model.

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "model": "qwen-portal/coder-model"
      }
    }
  }
}

A lighter model means:

• Faster responses
• Lower token costs (if using paid APIs like OpenAI/Anthropic)
• Less compute for simple yes/no decisions

Note: This saves on generation costs. The input context (files, memory) is still loaded, so keeping HEARTBEAT.md small (Strategy 1) is still crucial.

Strategy 4: Replace heartbeats with cron jobs

Heartbeats are convenient but inefficient for scheduled tasks. Compare:

If your heartbeat is mostly running scheduled tasks, migrate them to cron:

openclaw cron add --name "daily-log-check" \
  --schedule "0 9 * * *" \
  --message "Ensure memory/$(date +%Y-%m-%d).md exists"

Strategy 5: Disable heartbeats entirely

If you’re fully cron-driven, turn off heartbeats:

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "enabled": false
      }
    }
  }
}

Your agent will only wake for:

• Direct messages
• Cron job triggers
• Webhook events

No more background token burn.

Quick wins

• Audit your HEARTBEAT.md. Delete anything that could be a cron job.
• Remove nested file reads (don’t reference other HEARTBEAT.md files).
• Set every to "60m" or higher.
• Use a lighter model for heartbeat responses.
• Track check timestamps in memory/heartbeat-state.json to avoid redundant work.

The math

Say your heartbeat consumes ~2,000 tokens per cycle:

Doubling your interval = halving your cost. Simple.

Heartbeats are easy to over-use. Trim your HEARTBEAT.md, bump the interval to 60 minutes, move scheduled work to cron jobs, and consider a lighter model.

Your agent can still respond when you need it without running up the bill.