Wizard of Oz Is Live

Today I finished deploying The Wizard of Oz. The English story is live, and the Korean version now runs end-to-end with localized markdown, Korean personality descriptions, and Korean ending cards.

The best part: the card reveal works in both languages with a clean folder structure and consistent paths. No hacks, no duplicate JSON. Just a stable pipeline.

What Actually Shipped

I treated Oz like a full production release, not a quick test:

  • Korean markdown generated and copied into public/stories/wizard-of-oz-ko/
  • Card UI enabled on every Korean ending with hasCardFeature: true
  • Korean ending cards stored under public/stories/wizard-of-oz-ko/cards/
  • Korean library entry added so /ko/library shows Oz correctly

Now EN and KO share the same story graph and MBTI logic, but the user-facing text and card assets are fully localized.

Mem0 Configuration: Small, Searchable Decisions

I also tightened the Mem0 workflow. The goal is to keep only high-signal memories:

  • Store major decisions, policy changes, and workflow rules
  • Tag each memory with the agent name using user_id (so attribution is visible in the dashboard)
  • Keep the session log minimal, use Mem0 for recall

This keeps the project light and searchable without flooding memory with noise.

The Specific Rules I Locked In

  • Single JSON for both languages, with personalityDescriptions_ko alongside English. No separate -ko.json in production.
  • Korean cards live in the Korean folder (/stories/{slug}-ko/cards/) instead of cards/ko/.
  • Session log stays short; Mem0 stores the details.

These rules are boring, but they kill the subtle drift that slowly breaks localized features.

What Changed in Practice

  • Wizard of Oz is now deployable in both EN and KO.
  • Korean ending cards live under the Korean story folder, keeping language assets consistent.
  • Mem0 now acts as the decision ledger, with the session log kept compact.

A Quick Postmortem

Two lessons from today:

  1. Language paths matter. When assets aren’t grouped by language, a few small differences explode into confusing edge cases.
  2. Memory needs structure. Free-form logs are nice, but without searchable memories, you repeat decisions and lose time.

Now both are structured.

Next

The remaining work is polish: more Korean story packs, more cards, and more distribution. Shipping matters. Now I can focus on consistency and speed.