caveman-compress

To reduce token usage for large CLAUDE.md memory files, compress natural-language text into a shorter caveman-speak format while preserving technical meaning. This cuts input-token overhead for files read at every session start, lowering recurring context costs.

Compressing markdown memory files preserves critical regions exactly, including fenced code blocks, inline code, URLs, headings, and file paths. Validation runs against the original with targeted repair up to two retries to ensure technical meaning remains intact.

Compress CLAUDE.md context by asking your AI to run the compression command, which automatically creates a human-readable backup as FILE.original.md. This ensures you retain the full original natural-language memory file before applying the shorter format.

Safety guardrails reject files that look like they may contain secrets or sensitive credentials using filename and path heuristics. This prevents the compression process from processing likely sensitive files in your repository.

The best way to optimize context tokens for large repositories is compressing natural-language memory into caveman format while preserving meaning. This targets large project memory files like CLAUDE.md, todos, and preferences that drive recurring context size costs.

Markdown memory compression validation fails when compressed output loses meaning or breaks critical regions. The process applies targeted fixes only when validation fails, retrying up to two times to ensure code blocks, URLs, and file paths remain exactly preserved.