NovoMemo: AI Notes That Answer Only From What You Wrote

NovoMemo is a notes app you can ask questions. You write notes in markdown, then ask things like “what did I decide about the pricing tiers?” NovoMemo answers only from what you wrote and shows you exactly where it found the answer. If your notes don’t contain the answer, it says so instead of making one up.

I designed and built it on my own under my studio, Imaginovo. It’s live on the web with subscription billing, and the iOS and Android apps are on the way.

The problem

General-purpose AI chat is confident about everything, including things it doesn’t know. For personal notes that’s the wrong default: if I ask about my meeting, I want my notes or nothing. That set the requirements:

• Answers must come only from the user’s notes, never from the model’s general knowledge.
• Every answer must point back to its source passage.
• Asking must feel instant, and writing must work offline.
• It has to be a real product, with sign-in, billing, and usage limits, not a demo.

How retrieval works

1. Chunking. Each note is split at its markdown headings, and each chunk keeps the headings above it as context. Notes without headings fall back to roughly 1,000-character chunks with 200 characters of overlap, breaking on line ends where possible.
2. Embedding. Saving a note triggers an edge function that embeds its chunks with OpenAI text-embedding-3-small. Out-of-date chunks are re-embedded when a question comes in, so answers never use stale text.
3. Search. The question is embedded and matched by cosine similarity in Postgres with pgvector (HNSW index). No separate vector database is needed.
4. Relevance gate. If no chunk clears the similarity threshold, NovoMemo answers “not found” instead of handing the model nothing and hoping. This is the single most important design decision in the app.
5. Answer with citations. The model answers from the retrieved chunks only, streams the response back, and cites the chunks it used. Tapping a citation opens the note and highlights the passage.

Answers are routed through OpenRouter to the cheapest provider with a no-data-retention policy, with a fallback model when it isn’t configured.

Editing by chat

You can also ask NovoMemo to create or change notes. Proposed edits apply immediately with an Undo, and deletions always require confirmation. The step that classifies these requests never sees note text, which keeps the privacy boundary simple.

Architecture

• App: Expo (React Native + React Native Web) and TypeScript, with a CodeMirror 6 live-preview markdown editor.
• Offline-first: a local SQLite cache with a queue of pending changes that syncs to the server (last write wins).
• Backend: Supabase (Postgres, Auth with Google and Sign in with Apple, Deno edge functions). Business logic is shared between app and server and covered by Vitest tests.
• Billing: RevenueCat (Apple in-app purchase and web checkout, with Stripe processing cards). Webhooks are signature-checked and weekly usage limits are enforced on the server, not the client.
• Hosting: Cloudflare Pages for the web app and a Cloudflare Worker for the landing page.

How I built it

I built NovoMemo in about 12 weeks of part-time work, using AI coding agents (Claude Code) for most of the implementation. My job was the part agents can’t do on their own: deciding what to build and keeping it coherent. That meant a domain glossary, nine architecture decision records, written specs and tickets for every feature, and a test-first loop so each change was checked before it landed.

The lesson I’d bring to any team: AI makes writing code cheap, so the leverage moves to clear requirements, explicit decisions, and good tests.

Status

NovoMemo is live on the web at novomemo.com with free and premium plans. It launched recently, so I don’t have usage numbers to share yet.