caveman-compress

To reduce input token usage for large CLAUDE.md memory files, you can compress the natural-language Markdown into a shorter "caveman" style. This rewrite preserves technical substance, Markdown structure, and code blocks while shrinking the repeated per-session token cost and speeding up interactions.

Yes, markdown compression preserves code blocks, URLs, inline code, file paths, commands, and headings. It validates the output structure against the original file and performs targeted repair retries up to two times if any structural element is missing from the shortened text.

The best way to shorten todo lists and preferences without losing meaning is to apply a prose rewriting technique that trims natural language into a condensed style. The text is shortened while keeping all technical substance and Markdown formatting intact for repeated note-taking workflows.

You cannot compress context files that contain secrets, credentials, or API keys. The compression process includes safety guardrails that heuristically flag sensitive files and refuse to process them, preventing unsafe exfiltration of sensitive data during the token-saving rewrite.

When context optimization is applied, your original memory file is preserved through a non-destructive workflow. The system writes a backup as FILE.original.md before overwriting the target file with the compressed version, ensuring you can always revert to the full text.

Yes, there are file size limits for token saving on memory markdown files. The compression refuses to process empty files or large files exceeding 500KB, and it skips files heuristically flagged as sensitive to avoid unsafe operations and maintain processing efficiency.