Turbovec MCP: Embedded Hybrid Graph-Vector-Lexical Memory System

Turbovec MCP Server is an enterprise-grade, local-first Model Context Protocol (MCP) implementation that serves as an advanced, persistent, transaction-safe, and self-organizing long-term memory for AI coding assistants (such as Cursor, Claude Desktop, Cline / Roo Code, Windsurf, and Zoo Code).

By combining in-process SQLite graph storage, SQLite FTS5 full-text search, and the high-performance turbovec vector index, Turbovec MCP provides your AI assistant with the ability to "remember," "associate," "traverse," and "synthesize" complex knowledge webs across multiple chat sessions—completely bypassing standard token context window limits.

The Problem it Solves

1. Context Window Exhaustion: Pasting thousands of lines of code or entire folders into active chat threads is expensive, degrades reasoning, and eventually overflows token limits.

2. AI Amnesia (Stateless Sessions): When you close a chat window, your AI assistant forgets everything—architectural decisions, custom guidelines, preference notes, and session takeaways.

3. Keyword Search vs. Semantic Search: Normal keyword matching (like grep or CTRL+F) fails when queries do not match source text exactly. Turbovec resolves this by combining keyword, vector similarity, and relational graph context.

4. Information Decay & Fragmentation: Over time, stored memories become disconnected, orphaned, or stale. Turbovec introduces an autonomous background engine that continuously clusters, deduplicates, and prunes memories.

Related MCP server: Heimdall MCP Server

Architecture & Memory Flow

Turbovec operates fully in-process inside your local workspace. All operations are atomic, local, and private.

graph TD
    Client["Any MCP Client<br/>(Claude, Cursor, Cline, Windsurf, ...)"]
    Server["Turbovec MCP Server<br/>(FastMCP stdio)"]
    
    subgraph "Embedded Memory Engine (Local)"
        DB["SQLite Database File<br/>(memory.db)"]
        Index["Turbovec Index File<br/>(index.tvim)"]
        Model["SentenceTransformer<br/>(all-MiniLM-L6-v2)"]
    end

    subgraph "Core Memory Layers"
        Graph["Graph Engine<br/>(nodes, edges)"]
        FTS5["Lexical Search Table<br/>(entities_fts)"]
        Ontology["Ontology Manager<br/>(ontology.json Validation)"]
    end

    subgraph "Autonomous Background Engine"
        Librarian["Librarian Service<br/>(Clustering & Concept Synthesis)"]
        Radar["Semantic Radar<br/>(Relationship Discovery Daemon)"]
        Lifecycle["Lifecycle Manager<br/>(Archive & Pruning)"]
    end

    subgraph "Retrieval Pipeline"
        VectorChan["Vector Channel"]
        LexicalChan["Lexical Channel"]
        RRF["Reciprocal Rank Fusion (RRF)"]
        Reranker["Cross-Encoder Reranker (Optional)"]
        Telemetry["Telemetry Module<br/>(Search Stats Tracking)"]
    end

    Client <-->|"MCP (stdio)"| Server
    Server --> Model
    Server --> Graph
    Server --> FTS5
    Server --> Index
    
    Graph --- Ontology
    DB --- Graph
    DB --- FTS5
    
    Server --> Librarian
    Server --> Radar
    Server --> Lifecycle
    
    Server --> RetrievalPipeline["Hybrid Search Pipeline"]
    RetrievalPipeline --> VectorChan
    RetrievalPipeline --> LexicalChan
    VectorChan --> RRF
    LexicalChan --> RRF
    RRF --> Reranker
    Reranker --> Telemetry

For a detailed look at the storage schema, database tables, and metrics structures, see the Configuration Guide.

Quick Start (Local Setup)

Get Turbovec MCP up and running on your machine in under a minute:

1. Clone the Repository:

   git clone https://github.com/henny-bee/Turbovec-MCP-Server.git
   cd turbovec-mcp-server

2. Set Up a Virtual Environment & Install Dependencies:

   python -m venv venv
   
   # Windows:
   .\venv\Scripts\activate
   # Mac/Linux:
   source venv/bin/activate
   
   pip install -r requirements.txt

3. Verify Server Execution:

   python main.py

   You should see a clean start log in server.log. Raw console progress bars are suppressed to keep MCP stdio communication clean.

Documentation Map

To keep this manual organized and easy to digest, detailed reference materials and configuration steps have been separated into focused guides:

• Configuration & Setup Guide — Detailed environment variables, database schema explanation, developer dashboard, Docker-compose setup, and editor integrations (Claude Desktop, Cursor, Cline/Roo Code).

• MCP Tools Reference Manual — Complete categorized documentation for all 35 powerful MCP tools and prompts (CRUD, Hybrid Search, Temporal travel, Semantic Radar, Librarian, Bottles).

• Custom Instructions & Rules — Recommended custom instructions and system rules for your AI assistant to leverage Turbovec proactively.

Testing

The repository includes a comprehensive, 43-point unit and integration test suite using pytest to verify graph consistency, transactional rollbacks, hybrid search scoring, and background daemon workers.

1. Install Development Requirements:

   pip install -r requirements-dev.txt

2. Run Tests:

   pytest

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

Sponsored by