Project
MemoryHub
Content-agnostic engine over a markdown memory store with schema-validated writes and an MCP server.
- Role
- Author (spec-driven, built with AI agents)
- Repository
- github.com/jkarancs/MemoryHub
- Python
- Pydantic
- Typer
- FastMCP
- LanceDB
- pytest
A reusable Python engine over a markdown memory store with YAML frontmatter: schema profiles, Pydantic-validated guarded atomic writes, hybrid semantic search (BAAI/bge-m3 embeddings + LanceDB + full-text fallback), a Typer CLI, and a FastMCP server for agent access. Content lives in separate repos (private + public export). Notably, the engine itself was built by AI agents from written multi-phase specs - a concrete demonstration of the spec-driven agentic methodology.
Deployment (since 2026-07-18): the MCP server runs as a persistent background HTTP daemon (streamable HTTP on a loopback port) rather than a per-session stdio process. A launcher script starts it once, singleton-guarded (port bind + pid file), and keeps the embedding model warm in one shared process - agent sessions attach to the running daemon instead of each paying a ~45s model load and ~1.4 GB of RAM.