kimhjo

Judgment and domain expertise. AI can write the code, but it can't tell you what's worth building or why the architecture will hurt you in two years.

PlanetScale had the same issue before they killed their free tier entirely. The problem is these companies need a middle tier that's profitable, and $5-10/mo often isn't worth the support overhead for them.

For your use case, self-hosted Supabase on a $5 Fly.io or Railway instance might actually be cheaper long-term. You lose the managed DX but you get predictable costs. Neon is also worth looking at — their free tier is more generous on egress and they have a $19 launch tier that sits between free and full pro.

The WAL mode fix for concurrent MCP clients is clever.

One thing I'm curious about — how are you handling memory relevance decay? If I stored context about a project six months ago that's no longer accurate, the hybrid search might still surface it confidently. Do you have any mechanism for staleness detection or does the user have to manually curate old memories?

The single-file backend is a nice constraint to work within.

Curious how you're handling conversation history at scale though — storing it per-user in-memory works fine for a demo, but what happens when the cell restarts? Does the built-in database persist across restarts or does history reset?