PomPom-AI MCP Server

How to use

You connect your MCP client to the PomPom-AI server to access memory features. Use the memory tools to save important facts, recall the full memory context for any user, and perform semantic searches to locate relevant memories. Saving memories creates structured embeddings and metadata, while retrieval returns context for richer, personalized responses.

Practical usage patterns include: saving a memory after you share something important (it will be stored and later retrieved), fetching all memories for a complete context, and performing targeted searches to surface specific memories. The server exposes three core tools you can invoke via the MCP protocol: save_memory(text: str), get_all_memories(), and search_memories(query: str, limit: int = 3). When you save memory, the system extracts key facts, generates embeddings locally, and stores both vector data and metadata for fast retrieval. When you fetch memories, you can review all stored items or request a focused set of results based on a query.

How to install

Prerequisites: Python 3.12+, an OpenRouter API key for Claude 3.7 Sonnet, and a Supabase PostgreSQL database configured for your environment.

# 1. Clone the project and install the package in editable mode
git clone <your-repo-url>
cd pompom-ai
pip install -e .

# 2. Configure environment variables
# Copy the example and update credentials
cp .env.example .env

Next, set up the environment values shown in the example. You will provide your own credentials for the LLM provider and database, and you’ll enable the SSE transport on port 8051.

TRANSPORT=sse
HOST=0.0.0.0
PORT=8051
LLM_PROVIDER=openrouter
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_API_KEY=your-openrouter-api-key
LLM_CHOICE=anthropic/claude-3.7-sonnet
DATABASE_URL=your-supabase-postgresql-url

Finally, start the server and verify connectivity.

python src/main.py

To test connectivity from a client, run the provided test script.

.	est_server.ps1

Additional sections

Configuration details, security considerations, and troubleshooting follow the memory-service setup and run steps. The PomPom-AI server uses local embeddings with ChromaDB and PostgreSQL metadata, so data remains on your machine and is backed by encrypted storage. The printed example below shows how a client configures access to the memory MCP endpoint.

{
  "pompom-ai": {
    "url": "http://localhost:8051/sse"
  }
}

Environment variables to configure (descriptions provided for clarity): - TRANSPORT: transport mode, set to sse by default - HOST: server host address - PORT: server port (8051 in examples) - LLM_PROVIDER: OpenRouter or other provider compatible with Claude 3.7 Sonnet - LLM_BASE_URL: base API URL for the provider - LLM_API_KEY: key for API access - LLM_CHOICE: model/variant used for memory processing - DATABASE_URL: connection string for your Supabase PostgreSQL instance

{
  "name": "pompom_ai_http",
  "type": "http",
  "url": "http://localhost:8051/sse",
  "args": []
}

Memory tools exposed to Qodo AI are defined as functions with specific purposes: save_memory(text: str) stores important information, get_all_memories() retrieves all memories for context, and search_memories(query: str, limit: int = 3) performs semantic search. Memory data is stored in ChromaDB (vectors) and PostgreSQL (metadata) for fast, reliable access.

Maintenance tips include monitoring the local ChromaDB path (./chroma_db/), ensuring PostgreSQL connectivity, and updating API keys as needed. If the server won’t start, recheck the .env configuration and port exposure. If memory saving fails, verify the database connection. If search fails, restart the server to refresh the vector index.

Configuration examples

MCP connection example in JSON shows how to connect the PomPom-AI server for memory operations.

{
  "pompom-ai": {
    "url": "http://localhost:8051/sse"
  }
}

Security and privacy

All embeddings are generated locally with no external embedding API calls. PostgreSQL storage is encrypted, and the processing is local to your machine, ensuring your memory data remains under your control.

save_memory

Store information in long-term memory by processing text, extracting facts, generating embeddings, and saving to ChromaDB and PostgreSQL.

get_all_memories

Retrieve all memories for a default user, with pagination, to provide full context.

search_memories

Perform a semantic search over memories using an input query and return the most relevant results.