OpenClaw: cutting costs by 80%

One model for everything. By default OpenClaw uses one model (usually Sonnet) for the main tasks, heartbeats and subagents. Sonnet costs $3 per million input tokens.

A bloated bootstrap. Every message starts by loading all the workspace files. A typical workspace is 45+ KB (~15,000 tokens). That's $0.045 for the bootstrap alone.

No limit on context. If the workspace has grown to 100 KB, OpenClaw will load all 100 KB into every message.

Heartbeats on an expensive model. Every 3-4 hours. Each heartbeat = bootstrap + check + reply. On Sonnet that's $0.10-0.20, or $2-5 a day.

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5-20250929",
        "fallbacks": ["anthropic/claude-haiku-4-5-20251001"]
      },
      "models": {
        "anthropic/claude-haiku-4-5-20251001": { "alias": "haiku" },
        "anthropic/claude-sonnet-4-5-20250929": { "alias": "sonnet" }
      },
      "subagents": {
        "model": "anthropic/claude-haiku-4-5-20251001"
      }
    }
  }
}

What's happening:

• primary — the main model for chats (Sonnet, $3/M input)
• fallbacks — automatic switch to Haiku on a rate limit ($0.25/M input)
• subagents.model — subagents always use Haiku (12 times cheaper)

Subagents on Haiku are fine. They do parallel subtasks: finding files, checks, simple queries. Haiku 4.5 handles them perfectly.

Check the current size

du -sh ~/.openclaw/workspace/
# Typical result: 45-80 KB

wc -c ~/.openclaw/workspace/*.md

The DRY principle: don't duplicate

The most common problem is information repeated across several files:

• The project description — in SOUL.md, AGENTS.md and MEMORY.md
• The list of tools — in TOOLS.md and AGENTS.md
• Commands — in AGENTS.md and MEMORY.md

Rule: every fact should live in one place. Other files link to it.

MEMORY.md is the main source of bloat

It grows over time. After a month — 20-30 KB. The fix:

1. Create memory/archive-YYYY-MM-DD.md for old entries
2. Keep only what's current in MEMORY.md
3. Target: MEMORY.md < 5 KB

AGENTS.md, HEARTBEAT.md, SOUL.md

Delete unused sections, simplify the instructions for the cheap model, remove duplicates.

Installation

# Install Ollama
brew install ollama

# Download the model (~2 GB)
ollama pull llama3.2:3b

# Start it (or open Ollama.app)
ollama serve

Set up the provider in openclaw.json

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://127.0.0.1:11434/v1",
        "api": "openai-completions",
        "apiKey": "ollama-local",
        "models": [
          {
            "id": "llama3.2:3b",
            "name": "Llama 3.2 3B",
            "reasoning": false,
            "input": ["text"],
            "cost": {
              "input": 0, "output": 0,
              "cacheRead": 0, "cacheWrite": 0
            },
            "contextWindow": 131072,
            "maxTokens": 8192
          }
        ]
      }
    }
  }
}

Assign Ollama to heartbeats

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "model": "ollama/llama3.2:3b"
      }
    }
  }
}

Result: heartbeats run locally, without the API. Cost: $0. Latency: 2-5 sec.

One line in agents.defaults:

{
  "agents": {
    "defaults": {
      "bootstrapMaxChars": 12000
    }
  }
}

If the workspace grows beyond the limit, OpenClaw will trim the bootstrap context instead of loading all of it.

Recommended values:

• 12000 — aggressive savings (for a compressed workspace)
• 20000 — moderate (a good balance)
• 30000 — light (for a workspace with lots of instructions)

Add a block of savings instructions to SOUL.md:

## Token Efficiency

- Answer briefly. Don't repeat the context back.
- For heartbeats: if there's nothing to report → only HEARTBEAT_OK
- Don't retell files — use links.
- Group checks: one message instead of several.

Without this block the assistant may retell file contents (+500-1000 output tokens), reply at length to a heartbeat instead of HEARTBEAT_OK, and send 3 messages instead of one.

With Token Efficiency instructions the assistant generates 30-50% fewer output tokens.

Every OpenClaw session is saved to a .jsonl file. Over a month that's hundreds of megabytes. Create a script:

#!/usr/bin/env bash
# cleanup-sessions.sh
set -euo pipefail

SESSIONS_DIR="$HOME/.openclaw/agents/main/sessions"
ARCHIVE_DIR="$HOME/.openclaw/sessions-archive"
DAYS="${1:-7}"

[ ! -d "$SESSIONS_DIR" ] && echo "No sessions" && exit 1

TOTAL=$(find "$SESSIONS_DIR" -name '*.jsonl' | wc -l | tr -d ' ')
OLD=$(find "$SESSIONS_DIR" -name '*.jsonl' -mtime +"$DAYS" | wc -l | tr -d ' ')
SIZE_BEFORE=$(du -sh "$SESSIONS_DIR" | cut -f1)

echo "Sessions: $TOTAL ($SIZE_BEFORE), older than $DAYS days: $OLD"
[ "$OLD" -eq 0 ] && echo "Nothing to archive" && exit 0

mkdir -p "$ARCHIVE_DIR"
ARCHIVE="sessions-$(date +%Y%m%d).tar.gz"

find "$SESSIONS_DIR" -name '*.jsonl' -mtime +"$DAYS" -print0 | \
  tar czf "$ARCHIVE_DIR/$ARCHIVE" --null -T -
find "$SESSIONS_DIR" -name '*.jsonl' -mtime +"$DAYS" -delete

echo "Done: $(du -sh "$SESSIONS_DIR" | cut -f1), deleted: $OLD"

chmod +x cleanup-sessions.sh
./cleanup-sessions.sh 7

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5-20250929",
        "fallbacks": ["anthropic/claude-haiku-4-5-20251001"]
      },
      "models": {
        "anthropic/claude-haiku-4-5-20251001": { "alias": "haiku" },
        "anthropic/claude-sonnet-4-5-20250929": { "alias": "sonnet" }
      },
      "bootstrapMaxChars": 12000,
      "heartbeat": {
        "model": "ollama/llama3.2:3b"
      },
      "subagents": {
        "model": "anthropic/claude-haiku-4-5-20251001"
      }
    }
  },
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://127.0.0.1:11434/v1",
        "api": "openai-completions",
        "apiKey": "ollama-local",
        "models": [
          {
            "id": "llama3.2:3b",
            "name": "Llama 3.2 3B",
            "reasoning": false,
            "input": ["text"],
            "cost": {
              "input": 0, "output": 0,
              "cacheRead": 0, "cacheWrite": 0
            },
            "contextWindow": 131072,
            "maxTokens": 8192
          }
        ]
      }
    }
  }
}