Compress content by type — text, JSON, logs, tool output and code — through dedup, truncation and pattern sampling, with 40 to 90 percent typical savings, reversible in memory or through an explicit bounded store with atomic writes, integrity checks, TTL and size limits; usable as a library, a CLI or an MCP server. Use when a large payload must fit a context window before an agent reads it. Not for the agent-to-agent binary wire format — use agent-compression; not for repeated-structure hashing — use token-compressor; not for deciding the per-turn policy — use token-shaper.
