Reminisce: The AI Memory Companion

A voice-first AI companion that restores independence for seniors. Powered by Gemini 2.0, RAG, and ElevenLabs to remember conversations, manage tasks, and combat isolation

Comment

Inspiration

For millions of seniors, aging comes with a silent struggle: the fading of precious memories. We watched our own grandparents struggle to remember simple things, like a family visit or a daily pill, and saw how this loss of control led to anxiety and isolation.

We realized that while AI is advancing rapidly, it is often leaving behind the people who need it most. Most "Assistants" are too robotic, too fast, or require complex typing. We wanted to build something different: A companion, not a tool. An AI that doesn't just process commands, but reminisces with you, remembers your life story, and gently guides you through your day.

What it does

Reminisce is a voice-first memory companion designed specifically for seniors with early-stage memory loss.

  • It Listens & Speaks: Using a hyper-realistic, soothing voice, it converses naturally without the user ever needing to touch a keyboard.
  • It Remembers Everything: It stores long-term facts (e.g., "Grandson Tommy likes apple pie") and recalls them in future conversations to provide comfort.
  • It Proactively Cares: If a user asks for a reminder, Reminisce instantly schedules it and alerts them visually and audibly (e.g., "Time to take your blood pressure medication").
  • It Visualizes Memory: A simple, high-contrast "Upcoming Tasks" widget updates in real-time as the user speaks, providing immediate visual confirmation.

How we built it

We architected a high-performance RAG (Retrieval-Augmented Generation) pipeline to ensure low latency and high empathy.

  • The Brain: We utilized Google Gemini 2.0 Flash for its sub-second reasoning capabilities, allowing the AI to detect intents (like setting a reminder) instantly.
  • The Memory: We used Google Vertex AI to generate embeddings and stored them in a Pinecone vector database. This allows the AI to perform semantic searches over past conversations to retrieve relevant context.
  • The Voice (ElevenLabs Twist): We integrated the ElevenLabs API using the specific "Voice Design" feature to engineer a custom persona. We tuned the stability and similarity settings to create a voice that is slower-paced and deeper-pitched, optimized for elderly hearing ranges.
  • The Nervous System: The backend is built with FastAPI (Python) handling the orchestration, while the frontend is a responsive React/Vite application designed with high-contrast accessibility standards.

Challenges we ran into

  • Latency vs. Empathy: We initially struggled with the delay between the user speaking and the AI responding. Seniors lose focus quickly if the silence is too long. We solved this by upgrading to Gemini 2.0 Flash and optimizing our Pinecone queries, cutting response time significantly.
  • The "Robotic" Trap: Standard TTS voices sounded too cold. We spent hours tweaking the ElevenLabs voice settings to find the perfect balance of "authority" and "grandparent-like warmth."
  • Hallucinations: Early versions would invent reminders that didn't exist. We implemented a strict verification layer in the backend that parses the AI's output into structured JSON commands before executing any system actions.

Accomplishments that we're proud of

  • Zero-UI Interaction: We successfully built a flow where a senior can use the entire app just by speaking—no typing required.
  • The "Cheat" Widget: We built a real-time polling system that makes the "Upcoming Tasks" widget pop up instantly on the screen the moment the AI confirms a reminder. It feels like magic.
  • Making AI Accessible: We didn't just build a cool tech demo; we built a product that our own grandparents could actually use.

What we learned

  • Context is King: A bot is just a bot, but a bot that remembers your grandson's name is a friend. The RAG pipeline wasn't just a technical requirement; it was the emotional core of the product.
  • Speed is an Accessibility Feature: For seniors, a slow app is a broken app. Optimizing for speed was just as important as the UI design.

What's next for Reminisce: The AI Memory Companion

  • Visual Memory: We plan to integrate Gemini's Vision capabilities so users can hold up a photo and ask, "Who is this?", and the AI will identify family members.
  • Real-World Integration: Moving from our custom scheduler to full Google Calendar OAuth integration for family sharing.
  • Smart Home Hardware: Porting the software to a dedicated device (like a Raspberry Pi) so it can live in the living room as an always-on companion.

Built With

Share this project:

Updates