Kept — Family Memory for Alexa+

Kept is the family memory layer for Alexa+. It blocks allergens, matches answers to a child's reading level, and stops duplicate purchases, using what your home already knows.

Comment

Inspiration

Alexa+ knows the world, but it doesn't know your home. It doesn't know a 5-year-old has a tree-nut allergy, that an 8-year-old is in 3rd grade, or that you already own a spare HDMI cable. Kept is a family memory layer that fixes this.

What it does

Kept sits between a family's question and Alexa+'s answer and grounds that answer in household memory:

  1. Family Meal & Allergy Shield: a trending walnut-pesto recipe is blocked because Leo (age 5) has a tree-nut allergy. Kept offers a nut-free recipe that fits the family's Ninja air fryer.
  2. Age-Grounded Curiosity Shield: when Nidhi (age 8, 3rd grade) asks "Why do stars explode?", Kept replaces college-level jargon with a 3rd-grade balloon analogy and suggests a book not already on her shelf.
  3. Duplicate Spend & Cable Guard: before buying a 34.99 USD HDMI switch, Kept finds that the TV has open ports and an unopened cable is already in the drawer.

How we built it

  • A self-hosted MCP server (spec 2025-11-25, Streamable HTTP) built with the official MCP SDK in TypeScript. It exposes household-care tools such as verify_household_care and get_household_profile.
  • A Next.js companion UI that simulates the Alexa+ voice experience and calls the MCP server through /api/mcp.
  • A deterministic policy-guard package that checks household facts (allergens, reading level, owned devices) before any recommendation is shown.
  • Amazon Bedrock (Nova) integration for multimodal verification.

Challenges

  • Keeping the demo instant and reliable: we moved the UI to render results immediately instead of blocking on network calls.
  • Getting MCP protocol details right (JSON-RPC envelope, Streamable HTTP) and covering them with protocol tests.
  • Making the product honest: the assistant should never claim safety when it can't verify a fact.

What we learned

Personalization is most valuable when it is grounded in stored household facts rather than generic AI output, and an MCP server is a clean way to give an assistant that context.

Honest scope

The companion UI simulates the Alexa+ voice agent. The household data is seeded demo data, and the MCP server is real and runs from the repo. Household memory in the live demo is not persisted between sessions.

What's next

Persistent per-family memory, real Alexa+ account linking, and learning household facts from corrections.

Live demo: https://kept-alexa-mcp.vercel.app Code: https://github.com/reachmeghanashyam-coder/kept-alexa-mcp

Built With

Share this project:

Updates

Submission history