GrassDump

AI-powered personal capture system. Dump thoughts via Telegram, AI classifies them as todos or dumps, and you can search your memory semantically.

What works

• Telegram bot: send any text, AI classifies and saves it
• Agent loop: creates todos with due dates, dumps for loose thoughts, searches past notes
• Semantic search: vector embeddings via pgvector
• Commitment detection: daily scan flags stale todos with nudges
• Scheduled reminders: minute-by-minute due checks + daily digest at 8 AM user-local time
• tRPC API: today view, dump feed, search, create, complete, edit, flip type

Structure

apps/
  server/            # Bun entry point — just boots Hono
  client/            # React Native (Expo) — not started yet

packages/
  core/              # Shared Zod types + date parser
  db/                # Drizzle models + DAO layer (Neon/pgvector)
  ai/                # OpenAI agent, embeddings, commitment detection
  service/           # Business logic — single gate between API and DB/AI
  api/               # Hono + tRPC routes, Telegram webhook, cron jobs

Dependency flow

core → db → ai → service → api → server

API only imports from service. Service orchestrates db and ai. Nothing reaches across layers.

Tech stack

Layer	Tech
Runtime	Bun
Monorepo	Turborepo
API	Hono + tRPC
Database	PostgreSQL (Neon) + pgvector
ORM	Drizzle
AI	OpenAI (gpt-4o-mini + text-embedding-3-small)
Auth	Better Auth (not wired yet)
Frontend	React Native + Expo (not started)

Getting started

bun install
# Run from repo root (Bun only loads .env from CWD)
bun run apps/server/index.ts

Needs a .env at repo root with:

• DATABASE_URL — Neon connection string
• OPENAI_API_KEY
• TELEGRAM_BOT_TOKEN
• DEV_USER_TELEGRAM_ID — your Telegram user ID (dev auth)

Docs

• Spec: spec/SPEC.md
• Implementation plan: spec/IMPLEMENTATION.md