MindStash

I lost a genuinely good idea last year. Thought I'd write it down in a minute. Didn't. By evening it was completely gone.

I tried everything. Notion, Obsidian, Pocket, Apple Notes. The problem was never capturing. Every app does that. The problem is nothing ever gives things back. You save something and it disappears into an organised void.

So I built MindStash.

How it works: You type a thought (max 500 chars), hit save. AI categorises it into one of 12 categories, extracts urgency, intent, priority, time context — 10 signals from one sentence. No decisions from you. Then a chat agent lets you find anything in plain English.

Stack: FastAPI + Next.js 15 + Supabase + Claude (Anthropic) for the agent. Deployed free on Vercel + Railway.

3 things I learned building it:

• The 500-char limit is a product decision, not a technical one. Short = less friction = more saves = the whole product actually works. Every time I tested relaxing it, the UX got worse.

• Two AI models beat one. Categorisation fires on every save — I use a cheaper OpenAI-compatible API for that. The chat agent uses Claude directly. The split cuts costs significantly.

• If you'll query on a field, make it a real DB column from day one. I put AI signals in a JSONB blob early on and paid for it in 4 extra migrations.

What's next: Semantic search is already built (pgvector + embeddings). Browser extension. Mobile-first redesign.

Live and free at mindstashhq.space (no card needed)

Happy to answer anything about the build.