MCP

cortexmd is an open-source long-term memory and code-navigation brain for AI agents, exposed over the Model Context Protocol. It gives an agent a persistent memory that survives across sessions and a way to navigate a codebase by querying an index instead of reading whole files into context. Self-hosted, MIT licensed, pre-alpha.

The private version was a genuinely good tool that only worked for me: it was wired to my own setup, reading my personal synced Obsidian vault, in a way nobody else could pick up and run. Generalising it is what produced cortexmd’s brain-vault model: the tool owns its own brain and is its sole writer, and any user attaches their own vault read-only, untouched.

Part three of the cortexmd series. How a Rust tree-sitter indexer turns a repo into a SQLite symbol database, so an agent can navigate code by querying an index at roughly 60 tokens per result instead of reading whole files into context.

How cortexmd gives an AI agent a memory that behaves like one: eight kinds of memory, a hot/warm/cold lifecycle that promotes what you touch and lets the rest cool, hybrid recall that fuses keyword and meaning, a session-start wakeup, and a scheduled dream that quietly tidies the brain while nobody is looking.

A capable coding agent forgets everything when a session ends and re-reads the same files every time, burning tokens. This is the story of how a private homelab MCP server grew two answers to those problems and became cortexmd, the first post in a four-part series.