Digital Twin: shared AI memory and collaboration through MCP

The problem

Useful personal context often stays inside the chat where it was created. Switching to another AI client can mean explaining preferences again or losing a plan that should have been reusable.

Digital Twin stores memories and plans in a shared database and exposes them through MCP. The same information can be accessed from ChatGPT, Claude Code or another compatible client. It also lets connected users involve each other's Twins in a collaborative plan.

My contribution

My team and I won first prize with Digital Twin at the MCP Apps hackathon in Turin on 28 March 2026. The Buildathon took place at OGR Torino, hosted by Fonderia with Manufact among its sponsors. We took the idea from a personal memory layer to a working demonstration of Twin-to-Twin collaboration in a single day.

This was a collaborative hackathon project. The public repository documents the implementation, demo and first-place result. The case study focuses on how the product connects persistent context, agent tools and a user-facing workflow.

Engineering decisions

The conversational client is the main interface; a Next.js web app manages memories, connections, plans and settings. A TypeScript MCP server exposes memory search, agent queries, collaborative planning, notifications and calendar actions.

PostgreSQL stores the shared state and pgvector supports semantic memory retrieval. The planning engine collects contributions from specialist agents and connected friends' Twins, then synthesizes a plan and saves it for later retrieval. Collaboration follows connections accepted by the users.

The prototype also includes OAuth authentication and optional Google Calendar integration. Persistent memory makes access boundaries a central design concern: portability between clients does not imply making every user's context public.

What shipped

A working, open-source prototype combining an MCP server, management dashboard, persistent memory and collaborative planning. It won first prize at the MCP Apps hackathon in Turin, sponsored by Manufact.

The demonstration shows a plan being created with multiple agents and retrieved through a shared memory layer. The first-place result recognized the prototype built during the event.

Architecture

AI client

ChatGPT or another MCP-compatible client invokes the Twin's tools.

MCP server

Exposes memory, connections, planning, calendar and notification tools.

Shared memory

Stores users, memories and plans in PostgreSQL with pgvector search.

Planning engine

Collects agent contributions, synthesizes a plan and persists it.

TypeScript · MCP · Next.js · PostgreSQL & pgvector · Claude

Explore the project

Explore the project through its demo and available resources.

• Source code, demo & event result
• MCP Apps hackathon · Turin
• Planning implementation