Llama Memory

MCP Memory Service for llama-server with persistent history and semantic memory via Postgres + PGVector.

Note: This project is intended for local/demo use only. Do not expose it directly to the internet without additional hardening (HTTPS, proper auth, backups).

What it does

• Semantic memory: save and retrieve memories by meaning, not just keywords.
• Conversation bridge: LLM creates conversations automatically; memories are linked.
• Cross-session recall: ask "what did we talk about before?" and get accurate answers.
• MCP protocol: works directly with llama-server's built-in MCP support.

Requirements

• Python 3.11 (Miniconda recommended)
• PostgreSQL 16+ with PGVector extension
• llama-server with --jinja flag (required for tool calling)
• nomic-embed-text running on llama-server (port 8081 by default)

Installation

# Clone the repo
git clone https://github.com/noualit/llama-memory-local.git
cd llama-memory-local

# Create environment
conda create -n llama-memory python=3.11
conda activate llama-memory

# Install dependencies
pip install -e .

Configuration

Copy .env.example to .env and edit:

cp .env.example .env

Example:

# Database
DATABASE_URL="postgresql://postgres:yourpassword@localhost:5432/llamamem"

# Llama-server (LLM)
LLAMA_SERVER_BASE_URL="http://localhost:8080"

# Embedding model (nomic-embed-text via llama-server)
EMBEDDING_MODEL_URL="http://localhost:8081"

# Embedding model name (default: nomic-embed-text)
EMBEDDING_MODEL_NAME="nomic-embed-text"

# Service port
SERVICE_PORT=9001

Setup database

Create the database and run migrations:

psql -U postgres -c "CREATE DATABASE llamamem;"
alembic upgrade head

The application also ensures basic schema on startup for convenience.

Run the service

# Using the script
.\scripts\run_server.ps1

# Or directly
python -m uvicorn app.main:app --host 0.0.0.0 --port 9001

Connect to llama-server

Add to your llama-server MCP configuration:

{
  "mcpServers": {
    "llama-memory": {
      "url": "http://YOUR_SERVER_IP:9001/mcp"
    }
  }
}

The service must be reachable from llama-server. Use the actual IP, not localhost if they run on different machines.

MCP Tools

Tool	Description
create_conversation	Create a new conversation session
list_conversations	List conversations with memory count
get_conversation_history	Get all memories in a conversation
search_memories	Semantic search across all memories
save_memory	Store an important fact or decision

System prompt

You can:

• Fetch the recommended system prompt from the service:

  • GET /system-prompt → returns plain text.

• Or paste this minimal version into llama-server:

MEMORY WORKFLOW:
- At the start of each new conversation, call create_conversation with a short title.
- Use the conversation_id from create_conversation when calling save_memory.
- Before answering questions about past topics, call search_memories FIRST.
- When the user shares important information, save it with save_memory.
- If list_conversations has previous chats, check get_conversation_history for context.

Health check

curl http://localhost:9001/health

Returns DB status, embedding service status, and tool count.

Architecture

High-level structure:

• app/main.py — FastAPI app, lifespan, /system-prompt
• app/settings.py — Pydantic settings from .env
• app/clients/embeddings.py — Calls nomic-embed-text for vectors
• app/db/engine.py — asyncpg connection pool (singleton)
• app/db/schema.py — Auto-creates tables on startup
• app/mcp/endpoint.py — MCP protocol handlers, rate limiter
• app/mcp/tools/ — Individual tool implementations
• migrations/ — Alembic database migrations

Development

# Run tests
pytest tests/ -v

# Run with auto-reload
python -m uvicorn app.main:app --host 0.0.0.0 --port 9001 --reload

See CONTRIBUTING.md for contribution guidelines.

License

MIT (see LICENSE file).