MedMemory MCP

Current medication apps remind seniors after they forget. MedMemory predicts cognitive decline weeks before the first forgotten dose—turning Alexa+ into a proactive health intelligence platform

Comment

Inspiration

Medication non-adherence among seniors is a massive $100 billion healthcare crisis causing over 125,000 avoidable deaths annually[cite: 3, 5]. Traditional reminder apps fail because they react only after a dose is missed, completely missing the root cause: undetected, silent cognitive decline[cite: 3, 5]. MedMemory v3.0 turns Amazon Alexa+ into a proactive health intelligence platform[cite: 3].

What it does

MedMemory v3.0 introduces three core pillars:

  1. Cognitive Fingerprint Engine (The Trump Card): Passively monitors six clinical speech markers (speech rate, pause distribution, vocabulary richness, filler-word density, response latency, and syntactic complexity) across routine Alexa interactions, autonomously alerting caregivers via Amazon SNS when deviation exceeds clinical thresholds—days before a missed dose[cite: 4, 8].
  2. Two-Lane Authority Architecture: Built on the Model Context Protocol (MCP over Streamable HTTP), it separates probabilistic language parsing (Lane A, read-only) from deterministic safety-critical writes (Lane B), making double-dosing and LLM medical corruption structurally impossible[cite: 3, 7, 11].
  3. Bedrock Guardrails Safety Shield: Intercepts unverified or dangerous medical queries and serves secure fallback anchors[cite: 8].

How we built it

  • Integration Spec: Model Context Protocol (MCP) spec 2025-11-25+ over Streamable HTTP[cite: 1, 11, 12].
  • Compute: AWS Lambda Function URLs (Python 3.12, fastmcp)[cite: 1, 11, 12].
  • Intelligence Plane: Amazon Bedrock (Nova 2 Lite for language parsing & cognitive scoring, Nova Pro Multimodal for visual descriptor extraction, and Guardrails)[cite: 11, 12].
  • Safety Kernel & Persistence: Amazon DynamoDB with conditional writes and KMS encryption[cite: 11, 12].
  • Orchestration: Amazon EventBridge Scheduler (daily cognitive scans) & Amazon SNS for out-of-band caregiver alerts[cite: 11, 12].

Challenges we ran into

  • Handling fastmcp streamable HTTP transport requirements on AWS Lambda Function URLs without cold-start latency breaches[cite: 17].
  • Ensuring strict IAM isolation between read-only LLM parsing (Lane A) and write-capable database operations (Lane B)[cite: 11, 17].
  • Tuning Guardrails custom policies to process natural medication logs without false positives[cite: 17].

Accomplishments that we're proud of

  • Successfully architected a 100% serverless, HIPAA-ready path-to-compliance system fulfilling all four judging criteria (Technical Implementation, Design, Potential Impact, and Quality of Idea)[cite: 1, 3, 11, 16].
  • Packaged our core deviation math into a standalone, MIT-licensed open-source library (medmemory-cognitive-fingerprint)[cite: 12, 18].

What we learned

  • The immense power of separating probabilistic language models from deterministic database kernels to guarantee medical safety.

What's next for MedMemory v3.0

  • Expanding clinical trial validations and integrating multi-language support for diverse senior demographics.

Built With

Share this project:

Updates

Submission history