Simple Memory Extension MCP Server

An MCP server to extend the context window / memory of agents. Useful when coding big features or vibe coding and need to store/recall progress, key moments or changes or anything worth remembering. Simply ask the agent to store memories and recall whenever you need or ask the agent to fully manage its memory (through cursor rules for example) however it sees fit.

Usage

Starting the Server

npm install
npm start

Available Tools

Context Item Management

• store_context_item - Store a value with key in namespace
• retrieve_context_item_by_key - Get value by key
• delete_context_item - Delete key-value pair

Namespace Management

• create_namespace - Create new namespace
• delete_namespace - Delete namespace and all contents
• list_namespaces - List all namespaces
• list_context_item_keys - List keys in a namespace

Semantic Search

• retrieve_context_items_by_semantic_search - Find items by meaning

Semantic Search Implementation

1. Query converted to vector using E5 model
2. Text automatically split into chunks for better matching
3. Cosine similarity calculated between query and stored chunks
4. Results filtered by threshold and sorted by similarity
5. Top matches returned with full item values

Development

# Dev server
npm run dev

# Format code
npm run format

.env

# Path to SQLite database file
DB_PATH=./data/context.db

PORT=3000

# Use HTTP SSE or Stdio
USE_HTTP_SSE=true

# Logging Configuration: debug, info, warn, error
LOG_LEVEL=info

Semantic Search

This project includes semantic search capabilities using the E5 embedding model from Hugging Face. This allows you to find context items based on their meaning rather than just exact key matches.

Setup

The semantic search feature requires Python dependencies, but these should be automatically installed when you run: npm run start

Embedding Model

            We use the [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct)

Notes

Developed mostly while vibe coding, so don't expect much :D. But it works, and I found it helpful so w/e. Feel free to contribute or suggest improvements.