mdvault 0.2.0

CLI + MCP server for hybrid BM25 + semantic search over local Markdown vaults

mdvault

Your Markdown notes, searchable in Claude Code.

Index any folder of .md files — Obsidian vault, ~/.claude/ history, project docs — and search them in natural language, directly from Claude Code or the terminal. Zero infrastructure. One .db file.

What it does

Search combines FTS5 BM25 and 256-dim vectors, fused with RRF and re-ranked on 7 signals (term coverage, heading match, path match, etc).

Each chunk carries its document context as a prefix ([path > title > heading]) before indexing, following Anthropic's contextual retrieval approach.

Other things worth knowing:

• Parses Markdown links and [[wikilinks]], finds backlinks and similar files
• Optional query expansion via local Ollama LLM
• Memories decay over time, gain confidence when hit, and get auto-promoted from recurring queries
• potion-base-8M embeddings, CPU-only, ~30MB download
• Incremental indexing (SHA256 change detection, only reprocesses what changed)
• MCP server so Claude Code can search your vault directly
• Everything in one .db file (FTS5 index + vectors + link graph + metadata)

Quick Start

Paste install https://github.com/sderosiaux/mdvault in Claude Code — it will follow these steps.

1. Install

uv tool install mdvault

2. Index your notes

mdvault index ~/.claude/

3. Add MCP server to ~/.claude/mcp.json:

{
  "mcpServers": {
    "mdvault": { "command": "uvx", "args": ["mdvault", "serve"] }
  }
}

4. Restart Claude Code, then ask: "search my notes for kubernetes setup"

Keep the index fresh

Indexing is incremental (SHA256 change detection). Set up a cron to keep it updated:

# Every 30 minutes
(crontab -l; echo '*/30 * * * * uvx mdvault index ~/.claude/ 2>/dev/null') | crontab -

Other install methods

uvx mdvault --help              # run without installing
pipx install mdvault             # without uv

Example

$ mdvault search "memory system LLM"

[1] 0.983  .claude/projects/.../ae863d59.jsonl:70
### Dedicated Memory Platforms
- **Mem0**: Universal memory layer. $24M raised (YC-backed).
  41K GitHub stars, 13M+ PyPI downloads...

[2] 0.870  .claude/projects/.../agent-a581b10.jsonl:2
## The mapping: CPU, RAM, disk, I/O
Andrej Karpathy posted in October 2023 that LLMs should be
understood "not as a chatbot, but the kernel process of a new OS."

CLI Usage

# Index your notes (downloads ~30MB model on first run)
mdvault index ~/.claude/

# Incremental update (only changed/new/deleted files)
mdvault index ~/.claude/ --incremental

# Search
mdvault search "nginx reverse proxy config"
mdvault search "ssh tunnel" --top-k 10

# Search with query expansion (requires Ollama running locally)
mdvault search "ssh tunnel" --expand
mdvault search "ssh tunnel" --expand --expand-model qwen3:0.6b  # default model

# Related notes: links, backlinks, and semantically similar files
# (file path is relative to the vault root)
mdvault related path/to/note.md

# Stats (includes memory & query analytics)
mdvault stats

# Store a memory (searchable alongside your files)
mdvault remember "Kafka timeout is controlled by max.poll.interval.ms"

# List stored memories (with confidence, hits, decay)
mdvault memories

# Show knowledge gaps (recurring queries with poor results)
mdvault gaps

# Delete a memory
mdvault forget <memory-id>

# Custom DB location
mdvault index ~/.claude/ --db ~/vault.db
mdvault search "query" --db ~/vault.db

Claude Code integration (MCP)

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "mdvault": {
      "command": "uvx",
      "args": ["mdvault", "serve"],
      "env": {
        "VAULT_DB": "/absolute/path/to/vault.db"
      }
    }
  }
}

If VAULT_DB is omitted, defaults to ~/.local/share/mdvault/vault.db (Linux) or ~/Library/Application Support/mdvault/vault.db (macOS).

MCP tools exposed:

Tool	Description
search_vault	Hybrid BM25 + semantic search. Filter by vault, source, namespace
related_notes	Links, backlinks, and semantically similar files for a given note
store_memory	Store a memory (auto-chunked, searchable alongside files)
delete_memory	Delete memories by id or namespace

Then ask Claude things like "search my notes for how I configured SSH tunnels" or "what notes are related to my kubernetes setup?".

Search pipeline

Query
  ├── FTS5 BM25 search  → top-75 (NEAR bigrams + focused AND clause)
  └── Vector search     → top-75 nearest neighbors
          │
          ▼
    Reciprocal Rank Fusion  (k=15, BM25 weight 4×)
          │
          ▼
    Re-ranking (7 signals)
      ├── Cosine similarity (continuous, from vec distance)
      ├── Query term coverage (squared, bonuses at ≥80% and 100%)
      ├── First-chunk coverage (intro paragraph = topic signal)
      ├── Heading match (H2/H3 vs query terms)
      ├── Title match (H1 vs query terms)
      ├── Path match (filename + parent dirs vs query terms)
      └── Overview boost (about/intro pages with high coverage)
          │
          ▼
    Content-hash dedup → top-N results

Files are split on ##/### headings (max 400 words, 50-word overlap, small sections merged). Each chunk gets a context prefix ([path > title > heading]) before embedding and FTS indexing.

Query expansion (--expand) calls a local Ollama model (default: qwen3:0.6b) to generate a paragraph a relevant document might contain, then appends it to the original query for vector search. BM25 always uses the raw query. Pull the model with ollama pull qwen3:0.6b.

Memories

Memories are searchable alongside files. Three things happen behind the scenes:

Decay. Memories fade over 180 days (floor 0.1). Every search hit resets the clock. Search for something regularly and it stays relevant. Stop, and it drops in ranking.

Confidence. Base score depends on source (user=0.7, agent=0.5, promoted=0.3) plus a log hit boost (capped at +0.3). Memories that keep getting matched climb higher.

Auto-promotion. Every search is logged and clustered by embedding similarity (cosine > 0.85). When a cluster hits 5+ occurrences:

• Good results (avg score >= 0.3): the best result becomes a permanent memory
• Bad results (avg score < 0.15): a knowledge gap is recorded

Search query
  ├── query_log + query_vec (logged)
  ├── hit tracking on memory results
  └── every 20 queries:
        ├── cluster_recent_queries (embedding similarity)
        └── maybe_promote (crystallize or flag gap)

Run mdvault gaps to see what you keep searching for but can't find.

Tech Stack

Component	Library
Embeddings	model2vec (potion-base-8M)
Vector search	sqlite-vec
Full-text search	SQLite FTS5
MCP server	mcp (official Python SDK)
CLI	typer
Linter	ruff
Query expansion	Ollama (optional)

Limitations

• English-optimized: potion-base-8M is trained mostly on English. Semantic search degrades on other languages (BM25 keyword search still works)
• Markdown only (no PDF, DOCX)
• Exact vector search — scales to ~500k chunks on commodity hardware

Development

git clone https://github.com/sderosiaux/mdvault
cd mdvault
uv sync --dev
uv run pytest -q

Install pre-commit hooks (ruff lint + format):

uv run pre-commit install

License

MIT