Echoes: Preserve Memories, Celebrate Life
What Inspired Me
"If you lose your hands, you need prosthetics. If you lose your mind — you lose everything."
55 million people worldwide live with dementia. When I came across Margaret McCallion's story — diagnosed at 51, sent home from her career within months, asking herself "if I can't work, what am I going to do?" — I kept thinking about one specific cruelty: she still had every memory. Her childhood, her career, the people she loved. But no tool existed to help her capture those stories before the window closed.
Every existing AI tool in this space monitors decline, detects falls, alerts carers. None get ahead of it. Echoes is built to do something different — preserve the person before decline takes them.
What I Learned
Reminiscence therapy is already an evidence-based clinical practice used by dementia carers worldwide. Guided conversation about past memories reduces anxiety, improves mood, and strengthens identity. Echoes is a digital, always-available version of this — free, accessible, requiring nothing but a browser.
The hardest problem wasn't technical. It was tone. Every word in the interface had to feel safe for someone who may be frightened and confused. Getting Claude to ask exactly one question and wait — with genuine patience — took more iteration than the entire UI.
Architecture
Three core components:
1. Conversation Engine Claude acts as a gentle life story interviewer. The system prompt instructs it to ask one question at a time, follow threads naturally, never correct or argue, and cover key life themes: childhood, family, school, work, love, proudest moments, favourite places.
2. Narrative Generator On session end, all messages are passed to a second Claude API call which transforms the raw Q&A into a beautiful memoir-style prose narrative — not a transcript, but a flowing life story.
3. Story Archive Sessions are saved locally and compiled into a readable, exportable archive the family can keep forever.
Context Window Management
Claude has no memory between API calls. Building continuity required passing prior story context into each new session. The approximate context usage per call:
$$C = S_{prev} + Q_{history} + R_{system}$$
Where $S_{prev}$ is the prior story summary, $Q_{history}$ is the conversation history, and $R_{system}$ is the system prompt. Older sessions are progressively summarised to stay within token limits.
Safety Features
- Crisis detection — distress keywords trigger an immediate warm banner with the Alzheimer's Society helpline (0333 150 3456)
- AI transparency — every session opens: "I'm an AI — not a person, but I'm here to listen"
- No medical advice — Claude is hard-instructed to always defer to a doctor
- Themes tracker — visual progress showing which life areas have been covered
Challenges I Faced
Tone over technology. Dementia patients are a vulnerable population. A cold response, a moment of impatience, a confusing interface — any of these could cause real distress. Every design decision was filtered through one question: would this feel safe to someone who is frightened and confused?
The one-question constraint. Claude naturally wants to be thorough. Training it to ask exactly one question and wait — the way a good carer does — required careful prompt engineering across many test conversations.
The empty narrative problem. When sessions are too short, the narrative generator needs enough signal to work with. Solved by instructing Claude to write warmly about what it does know, and note that more sessions will enrich the story further.
What I'm Most Proud Of
The moment during testing when the conversation stopped feeling like a chatbot — and started feeling like someone who had all the time in the world and truly wanted to listen.
That's what Echoes is supposed to feel like.
Echoes is not a medical device and does not replace professional dementia care. If you or a loved one needs support, please contact the Alzheimer's Society: alzheimers.org.uk or 0333 150 3456.
Log in or sign up for Devpost to join the conversation.