Memory MCP Server
MCP server for providing context preservation to LLMs
Features
- Uses PostgreSQL as storage backend
- Supports vector search (requires pgvector)
- Uses all-minilm-l6-v2 for embedding vector generation
Key Dependencies
- rmcp MCP SDK package
- Tokio Asynchronous runtime
- Diesel PostgreSQL client
- Hyper HTTP server
- burn Model execution engine
Build Instructions
You need to place the model.onnx file of all-minilm-l6-v2 into the ./assets directory, and use the script provided by burn-onnx to upgrade the model to opset version 16.
If output verification is needed, you need to place the pytorch_model.bin of all-minilm-l6-v2 into the ./assets directory, and use the test.py script to create embedding outputs.
Build Modes
- Default build is CPU-only (faster build time, suitable for CI/CD and Docker):
cargo build --release
- Enable GPU inference support at compile time:
cargo build --release --features gpu
The runtime --gpu flag only takes effect when the binary is built with --features gpu.
Runtime Configuration
The server supports both CLI arguments and environment variables. The .env file is loaded automatically at startup.
Quick Start
cargo run --release -- \
--host 0.0.0.0 \
--port 9180 \
--max-search-results 20
Enable GPU at runtime (only valid when built with --features gpu):
cargo run --release --features gpu -- --gpu
Run database migrations only (do not start MCP server):
cargo run --release -- --migrate-only --db-url postgres://postgres:password@localhost/memory_mcp_db
CLI Arguments
Environment Variables
Example .env:
DATABASE_URL=postgres://postgres:password@localhost/memory_mcp_db
MAX_SEARCH_RESULTS=20
ALLOWED_HOSTS=localhost,127.0.0.1
ALLOWED_ORIGINS=
RUST_LOG=info
Note: when both CLI arguments and environment variables are provided, CLI arguments take precedence.
Third-Party Model Notice
This project uses model assets from sentence-transformers/all-MiniLM-L6-v2 on Hugging Face.
- Source: https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
- License: Apache License 2.0
- Local license copy:
assets/all-MiniLM-L6-v2-LICENSE.txt - Third-party notice details:
THIRD_PARTY_NOTICES.md
The Docker build process uses the repository model asset (assets/model.onnx, managed by Git LFS) during compile time.
The model data is embedded into the compiled binary, so runtime distribution does not require shipping model/tokenizer files under assets/.
Linux Binary Distribution Notice
This project publishes Linux x86_64 static binaries via GitHub Actions CI.
GitHub Releases currently contain a single Linux artifact built for x86_64-unknown-linux-musl.
When distributing the Linux release binary (for example via GitHub Releases), distributors should also provide:
- a copy of
LICENSEfor this project; - third-party attribution details in
THIRD_PARTY_NOTICES.md; - the model license copy in
assets/all-MiniLM-L6-v2-LICENSE.txt.
For this project, Linux release archives do not need to include model files from assets/ because the model is embedded in the binary.
MCP Server Features
Designed based on a unified core table structure to replace dynamic table creation which easily leads to Catalog bloat. The system currently contains two core tables:
documents: Stores metadata of document collections (categories).memory_items: Stores specific memory vector chunks.
Core Table Structure
documents table structure:
CREATE TABLE documents (
id bigserial PRIMARY KEY,
name TEXT NOT NULL UNIQUE,
description TEXT,
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);
memory_items table structure:
CREATE TABLE memory_items (
id bigserial PRIMARY KEY,
document_id BIGINT NOT NULL REFERENCES documents(id) ON DELETE CASCADE,
summary TEXT NOT NULL,
summary_embedding vector(384),
content TEXT NOT NULL,
content_embedding vector(384),
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);