OpenClaw Context Window Optimization — Reduce Costs and Improve

Why Context Window Matters in OpenClaw

Every message your agent processes consumes context window tokens — and you pay per token. As your workspace grows (larger MEMORY.md, more loaded skills, longer conversation history), context usage grows too. Optimizing this is one of the highest-leverage things you can do to reduce costs and improve agent performance.

What Goes Into OpenClaw's Context

Each session, OpenClaw loads:

• SOUL.md — your agent's identity and rules
• AGENTS.md — workflow rules and behavior guides
• MEMORY.md (main session only) — long-term memory
• TOOLS.md — tool configuration notes
• Loaded skills (SKILL.md files for active skills)
• Conversation history (recent messages)
• Any files explicitly read during the session

The sum of all these = your context. Bigger context = higher cost per response.

Optimization Strategy 1: Keep MEMORY.md Focused

MEMORY.md is loaded into every main session. A 5,000-line MEMORY.md is expensive. Keep it to the essentials:

# MEMORY.md - Curated Long-Term Memory

## Active Decisions (max 20 entries)
- [date]: Decided to use Railway for backend hosting
- [date]: Stripe monthly billing, not annual

## Key Context (max 10 entries)
- Primary LLM: Claude Sonnet for routine tasks, Opus for complex
- Deploy schedule: weekly on Thursdays

## ARCHIVE
→ See memory/2026-01.md, memory/2026-02.md

Archive historical entries to dated files. Main MEMORY.md stays under 200 lines.

Optimization Strategy 2: Trim SOUL.md and AGENTS.md

Audit your workspace files for redundancy. Rules you wrote 6 months ago that you never trigger should be removed or archived. Every line costs tokens on every session.

# Audit command:
wc -l ~/.openclaw/workspace/SOUL.md \
         ~/.openclaw/workspace/AGENTS.md \
         ~/.openclaw/workspace/MEMORY.md
# Target: each file under 200 lines

Optimization Strategy 3: Lazy-Load Skills

Don't load all skills for every session. Only load skills relevant to the current task:

# In AGENTS.md:
## Skill Loading
Load GitHub skill only when task involves: GitHub, PRs, code, repos
Load Stripe skill only when task involves: payments, billing, invoices
Load HubSpot skill only when task involves: CRM, leads, contacts

Optimization Strategy 4: Summarize Conversation History

For long cron sessions, instruct your agent to summarize its own output rather than keeping raw tool call results in context:

## Cron Session Rules (AGENTS.md)
After each major operation: summarize results in 1-2 sentences before proceeding.
Do not keep raw API responses in context — extract only the relevant data.

Optimization Strategy 5: Use Cheaper Models for Routine Tasks

Not every task needs your most capable model. Use model routing:

# Cron tasks: use claude-haiku or gpt-4o-mini
openclaw cron add \
  --name hex-daily-report \
  --model anthropic/claude-haiku-3-5 \
  --task "Post daily metrics digest"

Measuring Your Context Usage

openclaw session status --show-tokens
# Shows: total context tokens, cost estimate, breakdown by source

Ready to unlock this for your workflow? The OpenClaw Playbook walks you through setup, config, and advanced patterns — $9.99, one-time.