A hierarchical markdown memory system that keeps AI context windows lean by loading only the room you need. The MCP server exposes five tools over stdio: list palace rooms, read files, write session logs (with guards against overwriting canonical data), full-text search, and batch reads for multi-room queries. Each directory is a locus containing specific knowledge, with INDEX.md as the entry point and session logs that accumulate until consolidation merges them into canonical files. Optional Ed25519 signatures protect against prompt injection by tagging tool outputs as TRUSTED, DATA, or CRITICAL-DATA. Benchmarks show 52% fewer context lines loaded compared to flat memory systems. Comes with Claude Code skills for navigation, consolidation, and audit workflows. Designed for agents that need structured recall without loading an entire knowledge base into every conversation.
claude mcp add --transport stdio nano-nimbus-locus uvx locus