Mirror maintains a knowledge graph across all your conversations — people, patterns, tensions, fears, recurring themes. When you mention your coworker again three weeks later, it already knows the history. The AI doesn't do therapy-speak ("it sounds like you're feeling..."). It responds like someone who was actually paying attention. Tech stack: - Next.js 15 (App Router) + TypeScript - Supabase (Postgres + Auth + Storage) - GPT-4o with a structured few-shot prompt to kill the therapy-bot patterns - Knowledge graph built from session summaries (GPT-4o-mini extracts entities) - Web Push for morning perspective notifications (reads your graph, sends something specific) - PWA — installable, works offline for reading The hardest part wasn't the memory — it was the prompt. GPT-4o defaults to "I hear you, that sounds really hard" regardless of what you say. Took many iterations of few-shot examples to fix it.

I built Mirror because I was tired of re-explaining

my life to AI every single session.

The core idea: long-term memory via a knowledge graph.

Mirror remembers the people, patterns, and tensions you mention

— and connects them over time.

The hardest part was killing the therapy-bot tone. GPT-4o defaults

to "it sounds like you're feeling..." no matter what. Took dozens

of few-shot examples to break that pattern.

Happy to answer any questions — especially about the memory

architecture or the prompt engineering approach.

Mirror - Private reflection with long-term memory

Replies