caveman-compress

You can reduce input tokens by compressing markdown memory files into shorter phrasing while preserving headings, code blocks, URLs, and file paths. This lowers the context token cost of loading files like CLAUDE.md at every Claude session start.

Compress markdown context safely by applying targeted compression that explicitly preserves critical structure. It keeps headings, tables, lists, fenced code blocks, inline code, commands, and file paths unchanged while shortening surrounding natural-language text.

Markdown compression overwrites the original file but first creates a FILE.original.md backup. It validates the output and uses targeted repair retries, aborting and restoring the original file if validation repeatedly fails.

You should not compress files containing sensitive credentials. The compression process refuses files that look sensitive, including keys, credentials, and common secret filenames or paths, to avoid sending sensitive content to an external API boundary.

Yes, optimizing Claude memory context requires the anthropic and tiktoken dependencies. These libraries handle natural-language compression processing and validate token savings for markdown files loaded during repeated AI assistant sessions.