Product app
Wasp and React for authenticated capture, timelines, category views, and account workflows
In Build
Memoato
I am building Memoato as a low-friction personal memory system: capture a real-life note in natural language, preserve what was actually said, and turn the useful parts into structured history without forcing life into a spreadsheet first.
WaspReactPostgresHetzner
Personal software case study
The product bet
Most tracking tools make the user choose a category, form, unit, and exact structure before recording anything. That friction is especially expensive on busy days, which are often the days worth remembering. Memoato has to accept imperfect input quickly while keeping the raw record, derived facts, privacy choices, and later corrections understandable.
- Role
- Founder and hands-on engineer
- Period
- 2025 - Present
- Stage
- Live private beta, active development
Product loop
From public record to trusted input
- 01Capture a short natural-language note from the web app, an MCP client, or another trusted integration.
- 02Store the raw entry first, then extract structured facts with deterministic rules and a scoped AI fallback.
- 03Match confident facts to existing categories, keep uncertain material visible, and avoid silently inventing history.
- 04Turn accumulated entries into timelines, progress views, and useful recall without making dashboards the primary input model.
System shape
Simple boundaries, explicit sources of truth
Source of truth
PostgreSQL through Prisma for users, raw entries, structured events, categories, and timestamps
Ingestion
Raw-first API pipeline with deterministic parsing, optional model fallback, and asynchronous processing
Integrations
Revocable, hashed API keys scoped to raw-entry writes, plus a private MCP bridge
Privacy
Cloud, encrypted-cloud, and local-first product modes kept as explicit user choices
Operations
Hetzner deployment behind Caddy and systemd, with GitHub Actions release automation
Engineering judgment
Decisions that keep the product legible
Raw before derived
The original note is stored before parsing begins. Structured events can be corrected or rebuilt without erasing what the user actually recorded.
Confidence is a product concern
A plausible parse is not automatically a fact. Confident matches can become events; uncertain notes stay visible instead of being forced into the wrong category.
Categories are views
Goals, charts, and category pages remain useful, but capture starts with the user's words rather than requiring the user to navigate the data model.
Narrow integration access
MCP and automation tokens can write raw entries without receiving broad account access, and every key can be revoked or expired.
Current proof
What is already delivered
- Natural-language quick logging with raw-entry preservation and asynchronous fact processing
- Timeline, category, goal, chart, ordering, and human-readable daily summary views
- Deterministic parsing plus optional AI-assisted extraction with unmatched notes retained
- Scoped API keys and MCP ingestion for low-friction logging from trusted tools
- Email and Google authentication, privacy modes, PWA support, and Hetzner production deployment
Next proof
The next step is to make recall as useful as capture: improve correction flows, alias memory, provenance for derived facts, and summaries that help without pretending the system knows more than the underlying entries support.