2nd/10_Wiki/Topics/AI_and_ML/Procedural Narrative Generation.md

---
id: wiki-2026-0508-procedural-narrative-generation
title: Procedural Narrative Generation
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [PNG, Story Generation, Computational Narrative, Drama Management]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [narrative, generative-ai, games, llm-applications]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python / TypeScript
  framework: LLM (Claude/GPT) + planner / state machine
---

# Procedural Narrative Generation

## 매 한 줄
> **"매 algorithmic story authoring — symbolic planners 부터 LLM-driven emergent narrative 까지"**. 1970s TaleSpin / 1990s drama managers (Façade) 의 lineage; 2019 AI Dungeon 으로 LLM era 시작, 2024-2026 currently agent-based simulation (Smallville, Voyager, Genesis) 에서 narrative emerges from LLM agents interacting in worlds. Hybrid 가 winning: LLM creativity + symbolic constraints.

## 매 핵심

### 매 paradigm timeline
- **Story grammars (1970s-80s)**: Propp's morphology, Rumelhart, TaleSpin (Meehan 1977).
- **Planning-based (1990s-2010s)**: STRIPS planners author plot (IPOCL, MEXICA).
- **Drama management (2000s)**: Façade (Mateas & Stern 2005) reactive narrative.
- **Neural (2017-2020)**: GPT-2/3 generated stories, Plug-and-Play LMs.
- **AI Dungeon era (2019-)**: open-ended LLM-driven IF.
- **Agent simulation (2023-2026)**: Park's Smallville, Voyager, Genesis world model — believable agents whose interactions form narrative.

### 매 challenges
1. **Coherence**: long-range plot consistency (LLMs forget).
2. **Causality**: events must causally connect.
3. **Character agency**: characters with goals/personalities, not just templates.
4. **Author's vs character's intent**: meta-level narrative goals vs in-world goals.
5. **Tellability**: not every simulation is a good story.

### 매 modern hybrid stack
- **World model**: structured state (locations, items, NPCs, relationships).
- **LLM as narrator**: takes state + recent events → next beat.
- **Symbolic planner**: enforces causal constraints, beats arc.
- **Memory / retrieval**: vector DB of past events for consistency.
- **Drama manager**: monitors arc (rising/falling tension), nudges plot.

## 💻 패턴

### LLM beat generator with state
```python
def generate_beat(state, history, beat_target):
    prompt = f"""You are a narrator. Current state:
Characters: {state.characters}
Location: {state.location}
Inventory: {state.inventory}
Recent events: {history[-5:]}

Beat target: {beat_target}  # e.g., "rising tension", "reveal villain"
Generate the next narrative beat (2-3 sentences) and any state changes as JSON."""
    resp = claude.messages.create(
        model="claude-opus-4-7",
        max_tokens=400,
        messages=[{"role": "user", "content": prompt}],
    )
    return parse(resp.content[0].text)
```

### Story-arc drama manager (Freytag)
```python
ARC = ["exposition", "rising_action", "climax", "falling_action", "resolution"]

class DramaManager:
    def __init__(self, n_beats=20):
        # Map beat index to arc stage
        self.targets = ([ARC[0]] * 3 + [ARC[1]] * 8 + [ARC[2]] * 2
                        + [ARC[3]] * 4 + [ARC[4]] * 3)

    def next_target(self, beat_idx, tension_so_far):
        target = self.targets[beat_idx]
        if target == "rising_action" and tension_so_far < 0.3:
            return "rising_action_inject_conflict"
        return target
```

### Goal-based character agent
```python
class CharacterAgent:
    def __init__(self, name, traits, goals):
        self.name, self.traits, self.goals = name, traits, goals
        self.memory = []  # recent observations

    def choose_action(self, world_state):
        # LLM call: given personality + memory + state, pick action
        prompt = f"""You are {self.name}. Traits: {self.traits}.
Active goals: {self.goals}.
Recent: {self.memory[-10:]}.
World: {world_state}.
What do you do next? Output: action(target, args)."""
        return llm(prompt)
```

### Causal consistency check
```python
def verify_beat(beat, state):
    """Reject beats that violate world rules."""
    issues = []
    for change in beat.state_changes:
        if change.kind == "use_item" and change.item not in state.inventory:
            issues.append(f"{change.item} not in inventory")
        if change.kind == "location" and change.to not in state.adjacent_locs:
            issues.append(f"non-adjacent move {state.location}->{change.to}")
    return issues  # caller re-prompts on issues
```

### Memory-augmented retrieval (Smallville-style)
```python
def retrieve_relevant(query, memory_db, k=5):
    """Score = recency + importance + relevance (cosine)."""
    now = time.time()
    scored = []
    for m in memory_db:
        recency = 0.99 ** ((now - m.t) / 3600)
        relevance = cosine(embed(query), m.embedding)
        score = recency + m.importance + relevance
        scored.append((score, m))
    return [m for _, m in sorted(scored, reverse=True)[:k]]
```

### Branching narrative (interactive fiction)
```python
def turn(state, player_input):
    candidates = generate_n_continuations(state, player_input, n=3)
    chosen = drama_manager.pick(candidates, state.arc_position)
    state = apply(state, chosen)
    state.history.append(chosen)
    return state, chosen.text
```

### Constrained generation w/ JSON schema
```python
schema = {"type": "object", "properties": {
    "narration": {"type": "string"},
    "state_changes": {"type": "array", "items": {...}},
    "tension_delta": {"type": "number", "minimum": -1, "maximum": 1},
}, "required": ["narration", "state_changes"]}
# Use response_format=json_schema with Claude / GPT-5
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Short interactive story | Pure LLM w/ state in prompt |
| Long campaign (RPG) | LLM + structured world DB + retrieval |
| Strong authorial intent | Planner + LLM surface text |
| Emergent simulation | Agent-based (Smallville pattern) |
| Branching narrative game | Drama manager + tree-of-beats |

**기본값**: LLM beat-generator + structured world state + drama manager + retrieval memory.

## 🔗 Graph
- 부모: [[Computational-Narrative]] · [[Generative-AI]]
- 변형: [[Interactive-Fiction]] · [[Drama-Management]]
- Adjacent: [[Large-Language-Models]] · [[Planning]]

## 🤖 LLM 활용
**언제**: games (RPG, IF), creative writing tools, simulation, training data for narrative tasks.
**언제 X**: factual reporting, deterministic content (ads/legal), regulated medical/legal narration.

## ❌ 안티패턴
- **Pure LLM, no state**: forgets at chapter 3, introduces dead characters.
- **Planner-only, no LLM**: prose feels mechanical.
- **No drama curve**: flat tension → boring.
- **Unbounded creativity**: LLM invents items/locations with no consistency check.
- **Single-shot 100k-token story**: better as iterative beats with state.

## 🧪 검증 / 중복
- Verified (Park et al 2023 Generative Agents UIST, Mateas & Stern 2005, Riedl & Bulitko 2013 survey).
- 신뢰도 A (theory) / B (rapidly evolving practice).

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full PNG history + 2026 hybrid LLM/agent stack |