Kintsu

An AI companion that helps you understand your attachment patterns and communicate better in relationships. Built on attachment theory — the science of how we connect.

The name comes from kintsugi, the Japanese art of repairing broken pottery with gold. The idea: relationships aren't weakened by their cracks — they're made more beautiful when you understand and mend them.

What it does

• Attachment style awareness — identifies whether you lean anxious or avoidant and tailors guidance accordingly
• Knowledge-grounded conversations — responses are backed by RAG over attachment theory literature, not generic advice
• Persistent memory — remembers your relationship context, partner dynamics, and past conversations across sessions
• Memory processing pipeline — extracts and links episodic, semantic, and procedural memories with temporal validity
• Concept network — maps attachment theory concepts (triggers, behaviors, strategies) and their relationships

Tech stack

Layer	Tech
Frontend	React 19, TanStack Start (SSR), TanStack Router
Styling	Tailwind CSS v4, shadcn/ui
AI	Vercel AI SDK, OpenRouter (Kimi-2.5), OpenAI embeddings
Backend	Convex (reactive database, vector search, scheduled functions)
Infra	Cloudflare Workers via Alchemy
Monorepo	Bun workspaces

Project structure

kintsu/
├── apps/web/                  # React frontend
│   ├── src/components/chat/   # Chat interface
│   ├── src/components/ui/     # Component library
│   ├── src/lib/agents/        # AI agent definition & tools
│   ├── src/routes/            # Pages + API routes
│   └── src/routes/api/        # Chat streaming & transcription endpoints
├── packages/backend/          # Convex backend
│   └── convex/                # Schema, queries, mutations, actions
├── packages/infra/            # Cloudflare deployment config
└── scripts/                   # Book ingestion pipeline

Getting started

# Install dependencies
bun install

# Set up Convex backend
bun run dev:setup

# Copy env vars from packages/backend/.env.local to apps/web/.env
# You'll need: CONVEX_URL, OPENROUTER_API_KEY, OPENAI_API_KEY

# Start everything
bun run dev

Open http://localhost:3001.

Ingest knowledge base

# Place your source text in books/attached_content.txt
bun run ingest:book

This chunks the text, generates embeddings via OpenAI, and stores them in Convex for vector search.

How the AI works

The Kintsu agent has two tools:

1. queryKnowledge — vector search over attachment theory book chunks and concept graphs
2. recallMemory — vector search over user-specific memories from past conversations

Core memory blocks (user profile, partner info, relationship context, preferences) are always injected into context. Long-term memories are retrieved dynamically based on relevance.

A background memory processing pipeline extracts insights from conversations, deduplicates them, and builds a concept network linking attachment theory ideas.

Scripts

Command	Description
bun run dev	Start all services
bun run dev:web	Frontend only
bun run dev:server	Backend only
bun run check	Lint + format (oxlint/oxfmt)
bun run ingest:book	Ingest knowledge base

Deployment

cd apps/web
bun run alchemy dev     # Local preview
bun run deploy          # Deploy to Cloudflare
bun run destroy         # Tear down

License

MIT