Antigravity Project Memory (antigravity-project-memory)

A production-grade, zero-dependency community Model Context Protocol (MCP) server that empowers Antigravity agents with instant cross-conversation memory, architectural recall, and sub-2ms algorithmic retrieval.

💡 The Problem & The Solution

In modern agentic development environments like Antigravity, each agent conversation runs in an isolated context window. When an agent creates architecture decision records (ADRs), fixes a stubborn bug, or designs a multi-agent system in one conversation, that context is lost to future conversations unless manually re-pasted.

antigravity-project-memory bridges this gap. It operates locally on your machine, directly reading Antigravity's local SQLite summaries (conversation_summaries.db), dialogue logs (transcript.jsonl), and markdown artifacts (brain/<id>/*.md).

Why Antigravity Users Choose This Over Heavy RAG

Related MCP server: MemoryClaw

🏛️ System Architecture

flowchart TD
    Client["AI Client (Antigravity / Claude / Cursor)"] -- MCP stdio --> Server["MCP Server (server.py)"]
  
    subgraph Engine ["Algorithmic RAG Engine (<2ms)"]
        Parser["Dialogue & Artifact Parser"]
        BM25["BM25 Okapi Index (Lucene Smoothing)"]
        SymbolBoost["Code Symbol & Path Booster (1.0x - 3.0x)"]
        Recency["Exponential Recency Decay"]
        Stitcher["Chronological Passage Stitcher"]
    end
  
    subgraph AntigravityStorage ["Local Antigravity Data Store"]
        DB[("conversation_summaries.db")]
        Transcripts["brain/<id>/logs/transcript.jsonl"]
        Artifacts["brain/<id>/*.md (Specs, ADRs, Plans)"]
    end
  
    Server --> Engine
    Engine --> AntigravityStorage

🚀 Quickstart & Installation

Option 1: Run Instantly via uvx (Recommended)

No manual virtualenv management required. Run directly inside any MCP client:

uvx antigravity-project-memory

Option 2: Install via pip

pip install antigravity-project-memory

Then run:

antigravity-project-memory
# Or
python -m antigravity_project_memory

Option 3: Install from Source

git clone https://github.com/R-Mabasha/antigravity-project-memory.git
cd antigravity-project-memory
pip install -e .

🔌 Configuration with MCP Clients

1. Antigravity IDE / Antigravity CLI

Add to your project's .antigravity/mcp.json or global configuration (~/.gemini/antigravity/mcp.json):

{
  "mcpServers": {
    "antigravity-project-memory": {
      "command": "uvx",
      "args": ["antigravity-project-memory"],
      "env": {
        "ANTIGRAVITY_WORKSPACE": "${workspaceFolder}"
      }
    }
  }
}

Note: If ANTIGRAVITY_WORKSPACE is not specified, the server automatically defaults to the client's current working directory (os.getcwd()).

2. Claude Desktop (claude_desktop_config.json)

On Windows: %APPDATA%\Claude\claude_desktop_config.json On macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "antigravity-project-memory": {
      "command": "uvx",
      "args": ["antigravity-project-memory"],
      "env": {
        "ANTIGRAVITY_WORKSPACE": "C:\\path\\to\\your\\workspace"
      }
    }
  }
}

3. Cursor / Windsurf

Add to your project's .cursor/mcp.json:

{
  "mcpServers": {
    "project-memory": {
      "command": "python",
      "args": ["-m", "antigravity_project_memory"]
    }
  }
}

🛠️ MCP Tools Reference

The server exposes 4 high-value tools:

1. search_project_memory

Performs ranked algorithmic BM25 retrieval over past sessions and brain artifacts for the active workspace.

Parameters:

• query (str): Keywords, code symbols, filenames, or architectural concepts.

• workspace_path (Optional[str]): Workspace root. Defaults to ANTIGRAVITY_WORKSPACE or os.getcwd().

• top_k (int, default: 5): Maximum number of stitched passages to return.

• char_budget (int, default: 2500): Max characters per stitched passage.

Example Output:

# Antigravity Project Memory Search Results
- **Query**: `authentication jwt middleware`
- **Workspace**: `c:\my-project`
- **Matched**: 2 relevant passages (1.14 ms retrieval latency)

## 1. [DIALOGUE] Architecture Design - Turn 4
- **Relevance Score**: `7.82` (BM25: `4.12`, Symbol Boost: `1.95x`, Recency: `0.97`)
- **Conversation ID**: `0c523d8d-c315-4dfb-ba59-0653bfd295c3`
- **Created**: `2026-10-31T14:22:10Z`
- **Tools Used**: `run_command`, `write_file`

### User Request
How should we configure the JWT validation middleware?

### Assistant Response
We configured `JwtAuthGuard` in `src/auth/jwt.guard.ts` using RS256 with 15-minute token rotation.

2. list_project_conversations

Discovers all past Antigravity conversation threads belonging to the workspace.

Parameters:

• workspace_path (Optional[str]): Target workspace root.

• limit (int, default: 20): Maximum conversations to return.

Output: Returns a clean markdown table listing conversation title, UUID, last modified timestamp, step count, and execution status, followed by brief previews.

3. get_conversation_transcript

Extracts the clean, human-readable dialogue history of a past session.

Parameters:

• conversation_id (str): Conversation UUID.

• max_turns (int, default: 30): Max recent dialogue turns to retrieve.

Output: Structured dialogue turns (User requests, tools summarized, assistant replies) completely stripped of internal XML tags (<USER_REQUEST>, <ADDITIONAL_METADATA>, <thought>).

4. get_conversation_artifact

Reads and returns full markdown specifications, ADRs, task lists, or walkthroughs saved in a conversation's brain directory (brain/<id>/*.md).

Parameters:

• conversation_id (str): Conversation UUID.

• artifact_name (str): File name (e.g. walkthrough.md, task.md, or implementation_plan).

Output: Full markdown contents. If the file is not found, automatically lists all available artifacts for that conversation.

⚙️ Environment Variables

⚡ Performance Benchmarks

Measured on standard developer hardware (Windows 11 / AMD Ryzen 9 / Python 3.14):

🧪 Testing

Run unit tests via unittest or pytest:

python -m unittest discover -s tests -p "test_*.py"

All tests execute against live Antigravity workspaces and verify protocol-level tool invocation via FastMCP.

🤝 Contributing

Community contributions are warmly welcome! Whether you want to add new ranking heuristics, support other MCP clients, or optimize parsing:

1. Check out CONTRIBUTING.md for local development setup and guidelines.

2. Read our CODE_OF_CONDUCT.md to maintain a welcoming community.

3. Open a feature proposal or bug report via our Issue Templates.

⭐ Support the Project

If antigravity-project-memory saves you time or context in Antigravity, give this repository a star on GitHub! It helps other Antigravity engineers discover the tool.

⚠️ Disclaimer

This is an independent community open-source project designed for interoperability with local Model Context Protocol clients. It is not an official product of Google LLC or Alphabet Inc., and is not affiliated with, sponsored by, or endorsed by Google.

"Google", "Antigravity", and "Gemini" are trademarks of Google LLC. All product names, logos, and brands are property of their respective owners.

📄 License

MIT License. See LICENSE for details.