openclaw - 💡(How to fix) Fix [Feature]: Token optimization flags: compact tool descriptions and system prompt sections for effective use of local hosted models [1 participants]

Summary

Add experimental token-saving flags (compact tool descriptions, compact system prompt sections, and optional memory injection skip) to reduce context consumption for resource-constrained deployments.

Problem to solve

Users running OpenClaw with smaller language models (GPT-4, Sonnet, Haiku), local model backends (Ollama), or cost-sensitive deployments face significant constraints from token limits and costs. The current system prompt and tool descriptions are verbose and designed for large context windows, consuming 800-900+ tokens for fixed sections alone. Users need a way to reduce token consumption without losing tool functionality or manually editing/removing features. Currently, there's no built-in mechanism to intelligently compact the system prompt and tool descriptions, forcing users to either accept high token costs, manually strip features, or avoid using OpenClaw with resource-constrained models.

Proposed solution

Introduce three new experimental flags under agentDefaults.experimental:

1. compactToolDescriptions - Replaces verbose tool descriptions with short display summaries before sending to the model. Significantly reduces token usage while keeping all tools callable.

2. compactSystemPromptSections - Replaces verbose fixed system prompt sections (Documentation, CLI Quick Reference, Self-Update, Output Directives) with compact 1-2 line summaries. Saves approximately 800-900 tokens per run while maintaining all sections.

3. skipAutoMemoryInjection - Skips auto-injecting memory files (MEMORY.md) and the memory plugin prompt section into the system prompt. Users can still access memory on-demand via tools when this is enabled (e.g., instructions in AGENTS.md).

These flags are particularly useful for:

• Users running agents with smaller/local language models
• Cost-conscious deployments with large token-heavy interactions
• Scenarios where context window is the primary constraint

All flags default to false to preserve current behavior, and users can enable them individually in openclaw.config under agentDefaults.experimental.

Alternatives considered

No response

Impact

Affected users/systems:

• Users deploying OpenClaw with local model backends (Ollama, LLaMA, etc.)
• Self-hosted or on-premise deployments with limited infrastructure budgets
• High-volume agent deployments where per-token costs significantly impact viability

Severity: Annoying

• Users wanting to use OpenClaw with smaller/local models currently cannot, as verbose prompts and tool descriptions consume too much of the limited context window
• Cost-sensitive deployments are forced to use larger, more expensive models to maintain functionality

Frequency: Always

• Every agent run incurs the token cost of verbose system prompt sections (~800-900 tokens) and detailed tool descriptions
• Users with context-window-limited models are continuously blocked from adoption

Consequences:

• Inability to use OpenClaw with resource-constrained setups (local models, smaller cloudmodels)
• Increased operational costs for cost-conscious deployments
• Lost opportunities to serve users who prefer or require local/smaller model deployments
• Users must choose between OpenClaw capabilities or model/cost constraints, rather than having both

Evidence/examples

I am running the changes in my local fork of openclaw and seeing ~45% token savings without any visible impact on the functioning of the agent

All flags enabled simultaneously:

Baseline total:   9,405 tokens
  System prompt:  2,311 tokens
  Tool schemas:   7,094 tokens

Optimized total:  5,244 tokens
  System prompt:  1,571 tokens
  Tool schemas:   3,673 tokens
  ─────────────────────────────
  Total saved:    4,161 tokens  (44.2%)
    From prompt:    740 tokens
    From tools:   3,421 tokens
  + memory skip: ~200–500 tokens additional  (production only)

With all three flags: up to ~4,661 tokens saved (~47%) per agent turn.

Additional information

No response