What it does
Exposes a local Obsidian markdown vault as searchable, persistent memory for MCP-compatible AI clients. Implements a full modern information retrieval stack—hybrid BM25 + multilingual semantic embeddings fused via reciprocal rank fusion (RRF), BGE cross-encoder reranking, HNSW vector indexing with int8 quantization—all running locally with zero cloud calls. Additional features include late-chunking for improved context, PDF indexing with OCR support, and freshness-aware recall where each search result reports the note's age and last-modified timestamp.
Who it's for
AI developers, researchers, and knowledge workers building workflows across multiple LLM clients. Keeps Obsidian notes—where you already capture ideas, decisions, and research—accessible from Claude Code, Cursor, ChatGPT, and other MCP agents without vendor lock-in or session-to-session context loss.
Common use cases
- Query a personal knowledge base for project context while coding in Claude Code or Cursor
- Retrieve past design decisions or research conclusions across AI sessions
- Ground agent responses in note-age metadata to identify and re-verify stale facts
- Cross-reference markdown notes with BM25 full-text + semantic search during development
- Maintain vendor-neutral long-term memory using Obsidian files, not cloud-locked services
Setup pitfalls
- Requires filesystem read and write access to the Obsidian vault directory for indexing
- One credential detected in repository—audit or mask before use in shared or production environments
- Embedding models download from HuggingFace on first run, requiring internet connectivity for initialization
- Index size scales with vault size; HNSW + int8 quantization reduce memory overhead but monitor on constrained systems