Projects/active
Locus
Hierarchical Markdown memory for AI agents, shipped as an MCP server. Load only the room you need, not the whole palace.
Locus is a memory system for autonomous agents built on a simple idea: keep the context window small. Knowledge lives in a tree of Markdown files, a “palace”, where every directory is a room. An agent reads a short INDEX.md first, navigates to the one room that matters for the task, and reads only that. Session logs accumulate as append-only files until a consolidation pass folds them back into the canonical documents.
The name comes from the method of loci, the memorization technique that attaches facts to places in an imagined building. Each directory is one locus.
It ships on PyPI as locus-mcp, an MCP server any MCP-capable client can connect to, plus a locus init command that scaffolds a palace from a packaged example. Alongside the server are agent skills for navigating and consolidating a palace and audit tools for keeping it healthy. MIT licensed, Python 3.11 and up, published under the Nano-Nimbus organization.