mnemosyne-mcp

MCP server for Mnemosyne -- a 6-signal hybrid retrieval engine for code, documents, and database schemas. Reduces LLM context waste by 73%.

For the full reference, see MCP.md in the repository root.

Install

pip install mnemosyne-mcp

Register with Claude Code

claude mcp add mnemosyne -- mnemosyne-mcp

Or add to your project's .mcp.json:

{
  "mcpServers": {
    "mnemosyne": {
      "command": "mnemosyne-mcp",
      "args": []
    }
  }
}

Tools

search

Federated search across code and document partitions. Code results use 6-signal hybrid retrieval (BM25, TF-IDF, symbol matching, usage frequency, predictive prefetch, optional dense embeddings) fused via Reciprocal Rank Fusion. Document results use BM25 + TF-IDF with isolated vocabulary. Returns labeled sections so the LLM can perform cross-type ranking.

Parameters:

• query (string, required) -- natural language or keyword query
• budget (integer, default 8000) -- maximum token budget
• project_root (string, optional) -- path to project root

search_docs

Search the document partition only (PDFs, DOCX, CSVs, logs, and other non-code files). Uses BM25 and TF-IDF with an isolated vocabulary tuned for prose retrieval.

Parameters:

• query (string, required) -- natural language query
• budget (integer, default 8000) -- maximum token budget
• project_root (string, optional) -- path to project root

index

Index or re-index a codebase. Incremental by default (only processes changed files). Indexes both code and document partitions.

Parameters:

• project_root (string, optional) -- path to project root
• full (boolean, default false) -- force full re-index

stats

Show index statistics: file count, chunk count, tokens, language breakdown, chunk types.

Parameters:

• project_root (string, optional) -- path to project root

schema_ingest

Ingest database schema into the index. Accepts DDL files (.sql), JSON/YAML schema snapshots, or SQLite database paths for live introspection.

Parameters:

• source_path (string, required) -- path to schema source
• environment (string, optional) -- tag for the schema source (e.g., "production", "staging")
• project_root (string, optional) -- path to project root

schema_stats

Report indexed schema sources and statistics.

Parameters:

• project_root (string, optional) -- path to project root

How it works

Mnemosyne indexes your codebase and documents into local SQLite, scoring every chunk with retrieval signals fused through Reciprocal Rank Fusion. Code gets 6-signal hybrid search; documents get BM25 + TF-IDF with isolated vocabulary. AST-aware compression strips boilerplate while preserving function signatures, control flow, and documentation.

Zero runtime dependencies beyond Python 3.11+. No API keys. No cloud services. Everything runs locally.

License

AGPL-3.0 -- commercial licensing available from Cast Net Technology.