Inspiration
Caring for elderly family members experiencing memory loss is incredibly stressful. Caretakers often have to repeat the same daily facts (like when a visitor is arriving or where medication is kept), while seniors can feel frustrated or patronized when constantly asking for help. We wanted to build a bridge: a secure, local tool where caretakers can passively inject daily context, allowing the senior to ask Alexa natural questions and get warm, comforting answers.
What it does
Memory Aide is a Model Context Protocol (MCP) server that acts as a secure, local knowledge base for the Alexa+ track. It features a dual-interface:
The Caretaker Portal: A secure UI where family members or nurses can log daily facts (e.g., "Grandson Timmy is visiting at 2 PM", "Heart medication is in the top drawer").
The Alexa+ Simulator: The senior can ask natural questions. Alexa uses the MCP server to securely read the local facts and answer with empathy. Crucially, it includes AI Guardrails: If the user asks for sensitive information (like a bank password) or something not in the local database, the AI is prompted to safely deflect the question and gently remind the user to speak with their caretaker.
How we built it
Core Framework: We built the backend using Python and the FastMCP protocol to ensure strict compliance with the Alexa+ conversational track. AI Engine: The natural language generation is powered by Amazon Nova Micro via AWS Bedrock. We utilized the modern Bedrock Converse API to ensure lightning-fast, highly empathetic responses. Storage: We utilized a local, privacy-first JSON edge-storage system so that sensitive medical and personal facts never leave the user's home network until explicitly queried. Frontend: We built a custom FastAPI proxy and a sleek, Amazon-branded HTML/CSS web interface to simulate the end-to-end experience for the judges.
Challenges we ran into
Midway through the hackathon, we hit a major roadblock: our initial AWS Bedrock model (Claude 3 Haiku) was suddenly restricted as a Legacy model in our AWS account, throwing hard API exceptions. We had to pivot under pressure, completely rewriting our boto3 integration to dynamically query the AWS Bedrock API for active models, migrating our stack to the newer Amazon Nova model family using the Converse API. (Check out our friction_log.md for the full technical breakdown—we are claiming the 10% bonus!)
Accomplishments that we're proud of
We are incredibly proud of the AI's "bedside manner." By carefully tuning the system prompt and securely grounding it with local MCP data, we managed to get the Amazon Nova model to act with genuine empathy, gracefully deflecting harmful questions while contextually comforting the user.
What we learned
We learned the immense power of the Model Context Protocol (MCP). By separating the data layer (Caretaker facts) from the intelligence layer (AWS Bedrock), we realized how easy it is to build highly secure, privacy-first applications that still leverage world-class LLMs.
What's next for Memory Aide
Next, we want to integrate the Ring ecosystem! Imagine the Ring doorbell recognizing a face, querying the Memory Aide MCP server, and Alexa proactively announcing: "There is someone at the door. According to your caretaker, this is your grandson Timmy arriving for lunch!"
Log in or sign up for Devpost to join the conversation.