OpenClaw Token Optimizer

Overview

Deliver a practical audit and configuration plan that cuts input tokens and unnecessary calls while keeping answer quality. Provide concrete config edits, workspace file trimming guidance, and a prioritized rollout plan.

Workflow

1) Scope and locate configuration

• Identify the OpenClaw config file location (common paths include ~/.openclaw/openclaw.json, .openclaw/openclaw.json, or project root config).
• List injected workspace files in scope (e.g., AGENTS.md, SOUL.md, TOOLS.md, IDENTITY.md, USER.md, HEARTBEAT.md, MEMORY.md, and memory/YYYY-MM-DD.md).
• Confirm provider and model support for prompt caching and memory search to avoid proposing unsupported keys.

2) Baseline token sources

• Break input cost into buckets: system prompt, tool schema, workspace files, memory files, and conversation history.
• Use a rough sizing method if exact token counts are unavailable (e.g., characters/4 as a quick estimate) and call out that the estimate is approximate.

3) Input reduction (highest ROI)

• Trim workspace files first. Target budgets:
• AGENTS.md: keep only essential agent rules and policies.
• SOUL.md: reduce to short persona bullets.
• MEMORY.md: keep durable facts only; archive the rest.
• memory/YYYY-MM-DD.md: prune or rotate daily logs.
• Remove unused workspace injections in config (e.g., if TOOLS.md or IDENTITY.md is unused).
• Prefer memory search over full-file injection for large memories. If using qmd, index only needed paths.

4) Cache and context control

• Enable prompt caching for the primary model when supported. Set cacheRetention to a long window and keep a consistent system prompt to maximize cache hits.
• Configure heartbeat to keep the cache warm (e.g., ~55 minutes), using a low-cost model and a minimal heartbeat prompt.
• Enable context pruning with a TTL that matches the cache window to prevent unbounded history growth.
• Add compaction with memory flush so long sessions preserve durable decisions while clearing history.

5) Call reduction

• Audit cron and scheduled tasks. Consolidate overlapping checks, reduce frequency, and move non-creative tasks to cheaper models.
• Configure delivery to be on-demand or only on change to avoid no-op calls.

6) Model strategy

• Default to a cost-effective model for routine work and provide aliases for manual upgrades to premium models.
• Use subagents for parallel, isolated tasks with cheaper models to avoid bloating the main context.

7) Deliverables

Provide:

• A short audit summary and estimated savings.
• A concrete config patch or JSON snippet for openclaw.json.
• A list of files to trim, with before/after size targets.
• A phased rollout plan (quick wins first, then advanced options).

References

• Use references/openclaw-token-optimization.md for configuration snippets, checklists, and qmd guidance.