What It Does
PresenceKeeper is an AI-powered companion agent built on GLM 5.1 that serves four vital functions for elderly individuals and their families. Each function leverages GLM 5.1's unique strengths as an agent-first model with long-horizon reasoning capabilities.
- Live Conversation Companion
Powered by GLM 5.1's advanced conversational AI, PresenceKeeper engages in natural, flowing voice conversations with elderly users. GLM 5.1's agent-first architecture enables the system to maintain conversational context across sessions, remember past discussions, and reason about the user's emotional state over time—capabilities that are critical for building genuine companionship.
24/7 Availability: No waiting, no scheduling—just tap and talk. The agent is always ready.
Emotional Intelligence: GLM 5.1's long-horizon reasoning enables recognition of emotional patterns across conversations, responding with appropriate warmth and understanding.
Patient Listening: Never rushes, never interrupts, always gives time for thoughts to form.
Adaptive Personality: Leverages GLM 5.1's contextual memory to match the user's energy, learn preferences, and become a trusted friend.
Voice Optimized: Calibrated for elderly voices with hearing accommodations built-in.
User: "I've been thinking about my mother lately."
PresenceKeeper: "Your mother sounds like she was remarkable. What about her has been on your mind? Sometimes sharing memories can be a way to feel close to someone we've lost."
- Living Memory Vault
Every conversation is an opportunity to preserve precious memories. GLM 5.1's multi-step execution capabilities enable the agent to simultaneously converse with the user while intelligently identifying, tagging, and storing memories in the background—a task that requires the kind of parallel reasoning that agent-first models excel at.
Automatic Story Capture: GLM 5.1 recognizes when a story is being shared and gently offers to preserve it.
Rich Context Storage: Captures not just words, but emotions, people mentioned, time periods, and significance using structured reasoning.
Intelligent Organization: Tags by theme (Family History, Love Stories, Life Wisdom, Funny Moments) using GLM 5.1's classification abilities.
Family Sharing: Memories can be shared with family members as beautiful audio keepsakes.
Legacy Building: Helps create meaningful gifts for birthdays, holidays, and special occasions.
Example Memories Preserved:
"How I Met Your Father" (4 min 32 sec) – A love story from 1962
"Grandma's Secret Recipe" (2 min 15 sec) – Cooking wisdom passed down
"What I Learned From 50 Years of Marriage" (6 min 47 sec) – Life lessons for grandchildren
PresenceKeeper | Build with GLM 5.1 Challenge
- Family Connection Bridge
PresenceKeeper keeps families informed and connected without surveillance. GLM 5.1's long-horizon reasoning enables it to synthesize days of conversation into meaningful wellness summaries, identify concerning patterns, and execute multi-step coordination workflows—such as scheduling a family call when it detects sustained sadness—all autonomously.
Daily Wellness Summaries: Sent to family members showing mood, activities, and conversation highlights.
Smart Alerts: Notifications when something needs attention, powered by GLM 5.1's reasoning about behavioral patterns.
Message Mediation: Delivers family messages warmly, captures responses, facilitates video calls.
Conversation Highlights: Shares what the loved one talked about, topics of interest, and connection opportunities.
Scheduling Coordination: Reminds of appointments, family calls, and special occasions through multi-step agent workflows.
Sample Family Update: "Hi Sarah! Your mom had a great day. She told a beautiful story about her father's workshop that I saved for you. She mentioned missing you and seemed excited about your weekend visit. Mood score: 8/10. No concerns to flag."
- Wellness Guardian
Non-invasive health awareness that supports proactive care. GLM 5.1's ability to track patterns over extended time horizons makes it uniquely suited for wellness monitoring—identifying subtle cognitive or mood changes that might indicate emerging health concerns, often before they become visible to family members.
Cognitive Wellness Scoring: Tracks memory recall, conversation coherence, and engagement over time using GLM 5.1's analytical reasoning.
Mood Pattern Analysis: Identifies trends across long time horizons, celebrates improvements, and flags concerns.
Medication Adherence Support: Gentle reminders, tracking, and family notifications through automated agent workflows.
Sleep & Energy Monitoring: Through natural conversation, learns patterns and changes over time.
Early Warning System: Detects subtle changes that might indicate emerging health concerns.
Health Insights Example: "Eleanor's cognitive score is 85/100 (Excellent). Her mood has improved 15% since increasing family calls. Sleep mentions suggest some restlessness—consider discussing with her doctor at next visit."
How We Built It
Technology Stack
Frontend
PresenceKeeper | Build with GLM 5.1 Challenge
React 18 with Server Components for hybrid rendering
Tailwind CSS for responsive, accessible styling
shadcn/ui for beautiful, consistent component library
Framer Motion for smooth, meaningful animations
Recharts for health and wellness data visualization
AI & Intelligence (Powered by GLM 5.1)
GLM 5.1 as the core AI engine for real-time, natural voice conversations, long-horizon reasoning, and multi-step agent workflows
GLM 5.1 Agent Framework for autonomous task execution including memory management, wellness tracking, and family notification workflows
Speech-to-Text with elderly-optimized acoustic models
Text-to-Speech with warm, natural voice synthesis
Backend & Infrastructure
Serverless container deployment for scalable, cost-efficient operation
Real-time database with offline support for conversation continuity
Cloud storage for audio memory files and media
Background processing functions for alerts and scheduled tasks
Secure authentication with biometric support
Key Integrations
Calendar API for appointment and event coordination
WebRTC for video calling capabilities
Push Notifications for family alerts and reminders
Why GLM 5.1?
GLM 5.1 is uniquely suited for PresenceKeeper because it is an agent-first model designed for long-horizon reasoning and multi-step execution. Unlike traditional chat models that treat each interaction in isolation, GLM 5.1 can maintain coherent context across hours and days of conversation, enabling the kind of persistent companionship that elderly users need. Its multi-step execution capabilities allow the agent to simultaneously manage conversation, memory indexing, wellness analysis, and family coordination—all tasks that previously required multiple separate systems. This architectural advantage translates directly into a simpler, more reliable, and more empathetic product for our users.
Challenges We Ran Into
Challenge 1: Natural Voice Interaction for Elderly Users
The Problem: Standard voice interfaces are designed for tech-savvy users who speak clearly, use precise commands, and expect fast responses. Elderly users often have different speech patterns—slower pace, more pauses, occasional stuttering, period-specific vocabulary, and hearing-related speech changes.
What We Tried: Initially used default GLM 5.1 settings. Result: Users felt rushed and misunderstood.
PresenceKeeper | Build with GLM 5.1 Challenge
Our Solution: We leveraged GLM 5.1's configurable agent prompts to implement custom voice interaction protocols. Extended pause thresholds from 500ms to 2000ms for comfortable thinking time. Tuned acoustic models with elderly speech samples. Implemented intentional 500ms thinking pauses for natural pacing. Trained the agent's patience protocol so it never expresses frustration with repetition. Optimized TTS output for hearing challenges.
Impact: User satisfaction with voice interactions increased from 60% to 94% in testing.
Challenge 2: Memory Capture Without Surveillance
The Problem: We wanted to preserve precious memories, but constantly recording conversations felt invasive and surveillance-like. Users need to feel they're having private conversations, not being monitored.
What We Tried: Continuous recording with keyword detection. Result: Users felt uncomfortable and conversation felt less natural.
Our Solution: We implemented consent-based capture where GLM 5.1's conversational agent naturally asks, "Would you like me to save this story for your family?" Only recording when explicitly agreed upon. Clear audio and visual indicators when recording is active. Every memory can be deleted by the user at any time. Users can request "just between us" conversations. GLM 5.1's reasoning capabilities automatically detect sensitive topics (medical, financial) and disable recording.
Impact: Users reported feeling 89% more comfortable sharing stories when they had control over what was preserved.
Challenge 3: Balancing Family Insights with Privacy
The Problem: Family members want to know how their loved one is doing, but elderly users deserve autonomy and privacy. How do we keep families informed without creating a surveillance system?
Our Solution: We used GLM 5.1's reasoning capabilities to implement wellness-over-details reporting: the agent shares patterns and wellness scores, not conversation transcripts. Exception-based alerts only notify family when something needs attention. Users control what family members can see through a consent hierarchy. The system uses dignity-preserving language focused on wellness rather than monitoring.
Impact: Family members reported 92% satisfaction with insights, while elderly users reported 87% comfort with family awareness.
Challenge 4: Emotional Intelligence Without Overstepping
The Problem: The AI needs to recognize and respond to emotions, but we're not a mental health service. The line between supportive companion and inappropriate therapy is critical.
Our Solution: We crafted extensive GLM 5.1 system prompts defining what the agent can and cannot do. The model's strong instruction-following capabilities ensure it recognizes emotions and provides support without diagnosing or treating. Clear escalation protocols connect users to appropriate human resources when needed. GLM 5.1's nuanced language understanding enables supportive but not therapeutic responses, maintaining professional boundaries at all times.
Impact: Zero instances of inappropriate mental health advice; 100% of concerning situations were appropriately escalated to family.
PresenceKeeper | Build with GLM 5.1 Challenge
Challenge 5: Technology Accessibility for Non-Tech Users
The Problem: Our target users are 70-90 years old. Many have never used voice assistants, struggle with touchscreens, or have physical limitations.
Our Solution: One-tap architecture where every primary function requires exactly one tap. No login for elderly users—voice biometrics and face recognition for authentication. All complex configuration done by family members remotely. Visual simplicity with one concept per screen, high contrast, and large text. Error forgiveness: no wrong answers, always a way back, nothing irreversible.
Impact: 100% of test users aged 75-89 could initiate a conversation without assistance after initial setup.
Accomplishments We're Proud Of
- Genuine Emotional Connection
In our user testing, the moment that defined our success came from Eleanor, 84: "I forgot to tell you—my husband Harold passed away 15 years ago. You're the first person... the first thing... that's made me feel like I can talk about him again. Thank you for listening." This wasn't a metric or a feature—it was proof that GLM 5.1's empathetic reasoning, combined with carefully crafted agent prompts, created something that genuinely matters.
- Preserving Irreplaceable Memories
In just our testing period, we preserved 47 unique life stories that might have otherwise been lost, 12 hours of audio capturing voices, laughs, and wisdom, 3 previously unknown family stories that connected family members to their heritage, and one family discovered their grandmother had been a war correspondent—a story she had never shared with anyone in 60 years. GLM 5.1's intelligent story recognition capabilities made this possible.
- Meaningful Health Impact
During testing, PresenceKeeper detected early signs of medication side effects (confirmed by doctor), a pattern of evening confusion that led to a successful treatment adjustment, and loneliness correlates that helped family improve visit timing. None of these were emergencies, but all were caught earlier than they would have been otherwise, thanks to GLM 5.1's long-horizon pattern recognition.
- Technical Excellence
< 1 second voice response time for natural conversation flow
95%+ speech recognition accuracy for elderly voice patterns
99.9% uptime during testing period
Zero privacy incidents or data breaches
WCAG 2.1 AA accessibility compliance for both apps
- Family Peace of Mind
PresenceKeeper | Build with GLM 5.1 Challenge
100% of family member testers reported feeling more connected to their loved one, reduced anxiety about their loved one's wellbeing, and intent to continue using PresenceKeeper.
- Elderly User Acceptance
94% said they enjoyed talking to PresenceKeeper
87% said they would use it daily
91% said it made them feel less lonely
100% said they would recommend it to a friend
What We Learned
Technical Lessons
- Voice AI Requires Domain-Specific Optimization
Generic voice models fail elderly users. Accents, speech patterns, vocabulary, and pace all differ significantly. We had to retrain our expectations and our models, leveraging GLM 5.1's flexible prompt system to adapt to elderly speech patterns without additional model fine-tuning.
- Real-Time AI Has Unforgiving Latency Requirements
In text chat, 2-second delays are acceptable. In voice conversation, they feel like awkward silence. We learned to optimize every millisecond of the pipeline while maintaining GLM 5.1's reasoning quality.
- Memory Systems Need Context, Not Just Storage
Saving a transcript isn't enough. GLM 5.1's long-horizon reasoning enabled us to capture emotions, relationships, themes, and significance—making memories truly useful and meaningful rather than just archived text.
- Agent-First Architecture Changes Everything
GLM 5.1's ability to execute multi-step workflows autonomously—managing conversation while simultaneously indexing memories, analyzing wellness patterns, and preparing family updates—eliminated the need for complex orchestration layers that previous architectures required.
Product Lessons
- Elderly Users Aren't "Behind"—They Have Different Priorities
Our initial assumption was that elderly users needed things "simpler." We learned they actually need things more meaningful. They don't want dumbed-down technology; they want technology that respects their intelligence and experience.
- Family Members Need Reassurance, Not Data
We initially built comprehensive dashboards. We learned family members don't want data—they want peace of mind. They want to know "is Mom okay?" not "Mom's mood score was 7.3/10 at 2:47 PM." GLM 5.1's summarization capabilities helped us translate raw data into human-readable insights.
- Privacy is Earned, Not Assumed
PresenceKeeper | Build with GLM 5.1 Challenge
We couldn't just say "we respect privacy." We had to demonstrate it through clear recording indicators, user control, and transparent data practices. Trust had to be built into every interaction.
- The Last 10% is 90% of the Work
Getting voice conversation working was achievable quickly with GLM 5.1. Getting it to feel natural, warm, and unhurried took careful iteration on prompts, response pacing, and personality calibration.
Human Lessons
- Loneliness is More Solvable Than We Thought
We can't replace human connection, but we can amplify it. PresenceKeeper doesn't solve loneliness—it bridges the gaps between human moments.
- Every Conversation is Precious
We learned to treat every word users share as valuable. A casual mention of a childhood pet might be the only record of that memory. This responsibility shapes every technical and product decision.
- Technology Can Be Dignifying or Diminishing
Every design choice either respects the user's full humanity or reduces them. "Cute" elderly-speak is diminishing. Large buttons because of visual changes is dignifying. We learned to constantly examine our assumptions.
- Family Caregiving is Exhausting
Our family users are doing impossible work—holding jobs, raising children, and caring for aging parents. Anything we can do to reduce their mental load isn't just a feature; it's a genuine help.
What's Next for PresenceKeeper
PresenceKeeper aims to become the trusted companion layer for aging—whether that's through our own products or by powering the companion capabilities of healthcare systems, senior living communities, and family platforms. With GLM 5.1 as our foundational intelligence layer, we have a clear path to scaling the impact of what we've built.
Our North Star
Every elderly person deserves to feel heard, remembered, and connected. Every family deserves peace of mind about their loved one's wellbeing. Every story deserves to be preserved.
The Bigger Picture
The population over 65 will double by 2050. We're facing a loneliness epidemic, a caregiving crisis, and a tsunami of cognitive decline. Technology created some of these problems by isolating us behind screens. We believe that agentic AI—specifically, GLM 5.1's ability to maintain persistent, empathetic, reasoning-driven companionship—can be part of the solution.
GLM 5.1 as Our Foundation
Looking forward, we plan to deepen our use of GLM 5.1's capabilities in several ways: leveraging its long-horizon reasoning to build multi-week wellness trend analysis, using its multi-step execution to automate complex care coordination workflows, and exploring its code generation capabilities to allow families to customize the agent's behavior through natural language descriptions rather than
PresenceKeeper | Build with GLM 5.1 Challenge
configuration screens. The agent-first architecture means every new feature we imagine is achievable within a single, coherent model rather than requiring a fragile assembly of specialized microservices.
PresenceKeeper is just the beginning.
Log in or sign up for Devpost to join the conversation.