MemoryBridge

Inspiration

Challenge: For someone with early stage dementia or cognitive decline lose track of daily routines. Medication times and appointments slip out of memory. They need constant check-ins exahaust caregivers Most reminder apps just send notifications. They don't understand what caregivers actually need, they can't judge whether particular task is safe and they don't break instructions

What it does

MemoryBridge is multi-agent system that lets caregivers set remainders through natural conversations. The agents evaluate safety, simplify instructions, and send prompts that don't feel instructive or controlling

1. Understands the request - pulls out activity, time, and steps from natural language from caregivers
2. Rewrites it gently - turns the steps into short, one decision at a time instructions built for someone with cognitive decline
3. Waits for human approval - the caregiver sees exactly what the assisted person will see, nothing goes live until they approve it ## How we built it MemoryBridge is multi-agent system built on Strands Agents SDK, orchestrating four specialized agents: Routine Planning, Safety Reviewer, Friendly Communication, Escalation
4. Agents can draft, never activate: no agent in system ever holds a database write tool, every write happens in application code, triggered by caregiver action
5. Frontend: React, Next.js, split into caregiver console and Maria's example
6. Backend: A private FastAPI service hosting strands graph and local MCP subprocess
7. Cloud: Amazon Bedrock as model provider, all deployed on AWS Fargate behind separate public and private services. ## Accomplishments that we're proud of
8. Safety architecture where guarantees are enforced by code, not by prompting - agents are unable to write to the database and unsafe requests never reach a model
9. A complete working caregiver to approval assisted user loop, not just an isolated agent demo
10. A "Help-Me" flow that always reaches the caregiver, even if AI behind it falls or times out
11. A real deployment on AWS Fargate ## What's next for MemoryBridge
12. Support for more than one caregiver per assisted person, so whole family can stay in loop
13. Voice input, so a caregiver can speak instead of typing it
14. Letting a caregiver edit draft instead of approving or rejecting it

Built With

• amazon-web-services
• aws-fargate
• node.js
• react
• strands
• typescript