Local AI Memory Tool Helps Dementia Patients and Caregivers

When Rajamma steps outside and forgets where she is or why she left, her caregiver isn’t always by her side. A neighbor may want to help, but shouldn’t have access to private medical details. Later, when she visits the doctor, her caregiver might only recall fragmented memories. RememberMe CareGrid turns this chaotic moment into a structured care flow using local AI—specifically Gemma 4 E2B running through Ollama—to provide memory support, safety alerts, and coordinated community assistance without compromising privacy.

A Consent-Aware Memory Network for Dementia Care

Developed as a solo project for the Gemma 4 Challenge, RememberMe CareGrid is designed to address the gaps in traditional dementia care tools. Most existing solutions focus narrowly on GPS tracking or family-only reminders, leaving caregivers overwhelmed and patients confused. This prototype introduces a layered approach that respects privacy while restoring context and safety for all involved.

The system operates in three concentric circles of care:

• Patient layer: Provides memory cues through Lumo Companion, geofencing via SafePath to prevent wandering, and real-time responses to confusion. The goal is to help patients remember less alone.
• Family layer: Offers caregivers a unified dashboard with timelines, alerts, and doctor-ready summaries. The system ensures sensitive details remain accessible only to authorized users.
• Community layer: Enables trusted helpers—neighbors, ASHA workers, or volunteers—to assist with only the information they need, without exposure to private medical data.

This structure reflects the project’s guiding principle: Help patients remember less alone, and help caregivers care less alone.

Core Features Powered by Local AI

The prototype includes eight integrated components, all built around the patient’s immediate needs and the caregiver’s workflow:

• Lumo Companion: A patient-facing interface delivering calm memory cues and context when confusion arises.
• Trusted-person enrollment and recall: Recognizes caregivers and close contacts, ensuring only approved individuals receive access to relevant information.
• SafePath geofencing: Monitors safe zones and triggers alerts when a patient wanders beyond them.
• SOS escalation: Provides location and risk context to caregivers during emergencies.
• CareCircle task coordination: Assigns and tracks tasks among trusted helpers, ensuring no critical step is missed.
• CareLearn training cards: Generates practical, localized training resources using Gemma 4 to help caregivers and community workers respond effectively.
• Reward wallet: Encourages community participation through gamification, rewarding neighbors and volunteers for their involvement.
• Doctor Brief generation: Compiles structured, privacy-compliant summaries for medical visits, ensuring doctors receive clean, relevant timelines.

Each feature is designed to operate locally, minimizing data exposure and maximizing trust. The system never relies on cloud-based models for sensitive interactions, aligning with the need for privacy in dementia care.

How Gemma 4 Fuels the Care Workflow

Unlike many AI applications that use large models for decorative purposes, RememberMe CareGrid integrates Gemma 4 E2B as its reasoning core. The choice of a 31-billion-parameter dense model was intentional: it balances reasoning capability with the requirement for local, real-time responses in sensitive care scenarios.

The local pipeline follows a clear flow:

Patient input (voice or text)
  →  Local transcription or direct text input
  →  Gemma 4 E2B via Ollama for reasoning
  →  Structured validation and sanitization
  →  Output to dashboard, phone, or wearable device

Gemma 4 handles multiple critical functions:

• Generating calm, structured responses to patient questions like "Who am I?" or "Where am I?" using the care context.
• Producing reassuring safety cues when SafePath detects a wandering risk, rather than cold alerts.
• Creating practical training cards for neighbors or volunteers through CareLearn.
• Compiling doctor-ready summaries that prioritize clarity and relevance.

The system includes built-in safeguards: model outputs are validated against a strict schema, and malformed responses trigger fallback behaviors to prevent disruptions in care. For example, if an AI-generated response lacks required fields like intent or risk level, the app defaults to safer, predefined actions.

Why Local AI Matters for Dementia Care

Dementia care involves deeply personal and often stigmatized information—confusion episodes, family dynamics, medical routines, and safety risks. Sending every interaction to a remote server introduces unacceptable privacy risks. RememberMe CareGrid demonstrates that local AI can deliver meaningful assistance without sacrificing confidentiality.

The project’s architecture leverages Ollama to manage the local model efficiently, ensuring responses are both immediate and respectful of the patient’s dignity. By avoiding cloud dependency, the system also remains functional in areas with poor connectivity, a common challenge in caregiving environments.

This approach aligns with growing demand for privacy-preserving AI in healthcare. As wearable devices and personal assistants become more common in elder care, tools like RememberMe CareGrid show how technology can complement—not replace—human compassion.

Looking Ahead: From Prototype to Everyday Support

As a submission to the Gemma 4 Challenge, RememberMe CareGrid is still in its early stages. The demo reveals a promising vision: a care network that adapts to the patient’s reality, coordinates community support, and reduces the burden on exhausted caregivers. Future iterations could expand language support, integrate with wearable devices, or refine the AI’s ability to handle complex emotional contexts.

For now, the prototype stands as a testament to what’s possible when AI is deployed thoughtfully—prioritizing privacy, dignity, and real human needs over scale or speed. It’s not just another tool; it’s a step toward ensuring no one living with dementia has to face confusion alone.