Sophon

A deterministic context compression layer that runs as a single Rust binary with no ML dependencies at query time. Exposes 11 MCP tools including compress_prompt for structured input, compress_history for conversation memory, compress_output with 21 domain-aware filters for shell commands, and read_file_delta/write_file_delta for incremental edits. Built to stack in front of provider caching and memory systems like mem0, targeting the dynamic conversation blocks those layers don't touch. The real-data benchmarks show 84.7% token savings across a weighted blend of history, shell output, file reads, and search operations measured on the repo's own development cycle. Reach for this when you're burning tokens on repeated file reads, verbose tool output, or growing conversation windows in long-running agent sessions.