caveman-compress

You can reduce Claude memory token costs by compressing natural-language memory files like CLAUDE.md into a shorter caveman format. The process overwrites the original file with a compressed version while preserving code blocks, URLs, and headings exactly to maintain structure.

Caveman format compression is a token optimization technique that shortens natural-language memory files loaded at Claude session start. It reduces input tokens by rewriting text into a dense format while keeping code blocks, inline code, URLs, and file paths unchanged.

Yes, memory compression preserves markdown code blocks and URLs exactly. The validation process checks for mismatches and applies targeted fixes only if validation fails, ensuring inline code, file paths, and headings remain unchanged during token optimization.

To backup files before compressing Claude preferences, the skill automatically writes a human-editable FILE.original.md backup before overwriting the original file. This ensures you can always revert to the uncompressed natural-language version if needed.

Limitations of compressing markdown memory files include a 500KB file size limit to prevent accidental data exfiltration. Additionally, safety guardrails will reject files likely containing secrets or credentials to avoid compromising sensitive information during token optimization.

Yes, you can use tiktoken alongside anthropic dependencies for Claude Code context cost optimization. This combination enables the script to accurately measure token reduction when natural-language memory files are compressed into the caveman format.