EverAfter: AI-Powered Memory Preservation

EverAfter transforms personal memories into an interactive life graph using AI, semantic embeddings, and voice synthesis, preserving identity before memory fades.

Comment

Project Story – About EverAfter

Inspiration

Memory is fragile.

By 2050, over 150 million people are expected to live with dementia. When memory fades, families don’t just lose stories, they lose identity.

A father forgets his daughter’s wedding. A mother forgets her child’s voice. A son wishes he had recorded one more ordinary conversation.

We preserve photos. We preserve videos. But we do not preserve structured memory.

EverAfter was built to change that.

We asked a simple question:

What if we could model a human life as a semantic knowledge graph and preserve it before it fades?

What It Does

EverAfter is an AI-powered cognitive preservation platform that transforms personal memories into an interactive, explorable life archive.

Users can:

  • Record memories through text, voice, and images
  • Enhance them using AI
  • Explore their life as a connected memory graph

  • Ask questions like:

    • “When was I happiest?”
    • “How did I feel when I moved abroad?”

  • Hear responses narrated through neural voice synthesis

This is not a diary. It is a structured life model.

How We Built It

EverAfter is built as a cross-platform mobile application using:

  • React Native (Expo SDK 54)
  • Firebase (Auth, Firestore, Storage)
  • OpenAI GPT-4o-mini
  • text-embedding-3-small (1536-dimension embeddings)
  • Whisper (speech-to-text)
  • ElevenLabs (neural voice synthesis)

Intelligent Memory Enhancement

When a user enters a memory, we:

  1. Use structured prompt engineering with GPT-4o-mini
  2. Extract:
  • People
  • Locations
  • Emotions (16-class taxonomy)
  • Themes
  • Life stage classification
    1. Generate a concise narrative summary

If AI extraction fails, we use rule-based fallback heuristics to ensure 100% metadata completion.

Semantic Life Graph Construction

Each memory is converted into a 1536-dimensional embedding vector.

We batch embeddings in chunks of 50 to prevent API rate limits and reduce overhead.

To create meaningful connections:

  • We compute cosine similarity:

[ cos(theta) = {A x B}/{|A| |B|} ]

  • Apply a similarity threshold (≥ 0.82)
  • Select top-k connections to prevent O(n²) visual explosion
  • Limit edges per type to preserve clarity

We also cluster memories using a frequency-based person clustering algorithm with near O(n) complexity.

The result is a clean, explorable semantic memory galaxy.

Voice-Driven Memory Retrieval

When a user asks a question:

  1. We generate a query embedding
  2. Rank all memory embeddings using cosine similarity
  3. Retrieve top-3 most relevant memories
  4. Generate a first-person narrative response
  5. Convert it to natural speech using ElevenLabs

We implemented a finite-state audio machine:

IDLE → LOADING → PLAYING ↔ PAUSED → STOPPED

This ensures smooth playback and prevents UI blocking.

Challenges We Faced

  • Preventing visual clutter from O(n²) similarity connections
  • Managing embedding batch sizes to avoid API limits
  • Ensuring Firestore compatibility (removing undefined recursively)
  • Maintaining audio state synchronization
  • Designing emotionally powerful features without feeling manipulative

Balancing technical scalability with human sensitivity was our biggest challenge.

Why This Matters

EverAfter is powerful in real-world scenarios:

  • A father records advice for his child’s wedding and years later, his voice is still present.
  • An early-stage Alzheimer’s patient preserves their memories before cognitive decline progresses.
  • A family keeps ordinary voice notes that become priceless after unexpected loss.
  • A child in 2045 asks, “What were you like at 25?” and hears the answer directly.

We are not replacing people.

We are preserving what they chose to leave behind.

What We Learned

We learned that:

  • Technology should amplify humanity, not replace it.
  • Embeddings and clustering aren’t just technical tools, they can model identity.
  • Ordinary moments become extraordinary over time.
  • Presence through voice is more powerful than text alone.

What’s Next

  • Personalized voice cloning for premium users
  • Long-term memory trend analytics
  • Collaborative family memory spaces
  • Secure legacy release modes

The aging population crisis is growing.

The need to preserve identity is growing with it.

Technology already stores data.

EverAfter preserves a mind before it fades.

Built With

Share this project:

Updates

Submission history