NeuroStack
A local retrieval and maintenance layer for a Markdown knowledge base such as Obsidian or Logseq, indexing notes into SQLite with full-text search, embeddings and a link graph, and exposing it to any MCP client.
NeuroStack indexes a folder of Markdown files into one SQLite database combining full-text search, embeddings and a wiki-link graph. An MCP client queries it to search the base, walk the link graph and save memories; the query itself returns ranked evidence with note paths, and NeuroStack does not run a model itself, leaving the reasoning to the connected AI client.
Between queries it maintains the knowledge base: it flags notes that have gone stale, harvests decisions and root causes from AI sessions, merges recurring memories into learnings, and queues proven ones to become notes. Indexing itself never modifies the underlying files, though four opt-in MCP write tools can create, edit, delete or list notes, with every write committed and pushed through the vault's own git history and paths outside the vault or hidden directories rejected.
It installs from npm, works with Obsidian, Logseq, Notion exports or plain Markdown folders, and is intended to run locally with nothing leaving the machine unless a third-party LLM provider is configured. It is written in Python under the Apache-2.0 license.