Roadmap v3 — Emergence
The long-term vision. When the world starts surprising us.
Status
🔮 Vision — Target: TBD (after v2 stabilizes)
Goal
Reach the point where the simulation produces behavior that genuinely surprises its creators. Not scripted drama. Emergent complexity from simple rules + memory + LLM-generated text.
The North Star
"The most interesting thing about an AI world is not what you make it do. It's what it does when you're not looking."
v3 is the realization of this.
We stop being the authors. We become the maintainers of a system that writes its own stories.
Pillars
1. Generative Everything
v3 NPCs don't just act and write diaries. They:
- 🎨 Create art — paint, write poems, compose songs
- 📚 Write books — long-form text generated over months
- 🎵 Form bands — collaborative music creation
- 📰 Start newsletters — town-wide gossip sheets
- 🏛️ Propose laws — emergent governance
- ⚔️ Start feuds — long-running conflicts with backstories
2. Long-Arc Storylines
Stories that unfold over months of in-game time:
- 💔 A slow-burning romance over 90 days
- 🕵️ A mystery that takes 200 days to resolve
- 🏛️ A political crisis that splits the town
- 🎉 A cultural movement that sweeps through
These are not scripted. They emerge from accumulated decisions and memories.
3. Memory Decay & Consolidation
NPCs genuinely forget over time. Important things survive. Trivia fades.
This means:
- Old feuds can be forgotten (or remembered by grandchildren)
- Childhood trauma can resurface in adulthood
- Cultural myths form from distorted memories
4. Generational NPCs
NPCs have children. Children grow up. Grandparents die.
- Children inherit personality traits (with mutation)
- Children are raised by their NPC parents (in-game)
- Family trees span 50+ years of in-game time
5. Procedural Towns
The world can spawn new towns procedurally:
- New geography
- New NPC archetypes
- New cultural norms
- New event distributions
Each procedurally-generated town has its own "personality."
Technical Challenges
| Challenge | Approach |
|---|---|
| Memory explosion | Aggressive decay + summarization |
| LLM cost explosion | Caching + batching + smaller models for routine tasks |
| Coherence across long arcs | Periodic "narrative memory" consolidation |
| Procedural quality | Templates + LLM refinement |
| Performance at scale | Tiered compute (cheap rules for routine, LLM only when needed) |
What does NOT change
- ❌ No god-mode
- ❌ No combat
- ❌ No PvP
- ❌ No pay-to-win
- ❌ No ads
- ❌ No tracking beyond what the docs site needs
The Real Question
By v3, we want to answer:
Can a sufficiently rich simulation of LLM-driven agents, with persistent memory and emergent interaction, produce behavior indistinguishable from a small real community?
We don't know. We're going to find out.
Open Research Questions
- 🧠 How do NPCs form a stable identity across years of memory turnover?
- 💬 Can NPC conversations be coherent enough to drive multi-day narratives?
- 🤝 How do factions emerge and dissolve without scripting?
- 🎭 Can the LLM produce genuinely surprising (not just plausible) behavior?
- 📖 Will players still find it interesting after 100 hours of observation?
What we'd love to see (but won't promise)
- 🌐 Open data export — let researchers study the world
- 📚 Academic papers on emergent behavior in agent simulations
- 🎓 Educational use in universities (sociology, AI, game design)
- 🎮 Community-run sister worlds (separate instances)
What success looks like
Success at v3 is when:
- We read a diary entry and learn something we didn't know about the world.
- Players discover things we didn't design.
- The world continues to be interesting after a year of continuous operation.
When do we ship v3?
When v2 has been running stably for 6 months, and the technical foundations (memory, decay, LLM batching) are proven.
No rush. This is a research project as much as a product.