AI Dungeon Platform Research

1. Introduction

1.1 Developer and Launch

AI Dungeon is the earliest platform in this group (2019) and the originator of the "model as narrative engine" product form. It predates NovelAI by two years, 阅文妙笔 by four years, and 番茄 AI 工具箱 by five years.

1.2 Product Positioning

AI Dungeon is positioned as interactive fiction / AI Dungeon Master — players are not authors, but players-and-authors:

This difference determines AI Dungeon's Harness design focus: it invests in L2 (multi-style models) and L3 (open loop), rather than L4 (long-range state). In interactive fiction, a play session typically ends within a few hours; "remembering turn 1 at turn 500" is not a core requirement.

The arrival of Voyage changed this assessment — Latitude used World Engine to fill in the deterministic mechanics layer, showing they found "long-range state" to be the real ceiling bottleneck (see sections 3.4 and 4.3).

1.3 Model Lineup

2026 model lineup (fine-tuned by play style):

"Fine-tuned by play style" is the most distinctive design of this platform's L2 layer. It does not offer "stronger/faster" models, but models with "different personalities" — choose Wayfarer for brutal combat, Madness for chaos, Hearthfire for a daily-life atmosphere.

In the Harness context, this approach belongs to model routing in the L2 tools layer: giving the model-selection choice to users and routing by task type (here, play style) rather than by performance metrics.

Worth noting: the free tier's Mistral fine-tune has been publicly released on Hugging Face. This is quite rare among commercial platforms, and also explains why the community can understand its memory mechanism relatively deeply.

1.4 Pricing

Supplementary rules:

• Payable monthly / semi-annually / annually with discounts; Journey and Legend include a one-week trial.
• Shadow Tiers offer higher context and access to stronger models.
• Context window: paid tiers 4,000 to 32,000 tokens.
• Image credits: Adventurer 480, Champion 760, Legend 1,650, Mythic 2,750.

Memories is the only L4 capability on this platform directly tied to the subscription tier: from 25 at the free tier to 800 at Ultimate, a 32× difference. This is steeper than NovelAI's Lorebook capacity tiers (200 → 2,048 tokens, 10×).

1.5 Openness

2. Glossary of Terms

3. Feature Description

3.1 Core Features

3.2 Story Cards and Memory Banks

Story Cards + Memory Banks: store contextually relevant information, retrieved only when needed, maintaining narrative coherence over long sessions.

The two are the two instances of this group's "long-range state management Paradigm B · keyword-triggered entry library". AI Dungeon calls it "Story Cards", NovelAI calls it "Lorebook", and the mechanisms are nearly identical — indicating this paradigm has been validated twice in the industry.

The difference lies in how capacity is measured: NovelAI counts by token, AI Dungeon counts by number of entries (Memories 25–800). Counting by entries is more intuitive for players, but the length of a single entry cannot be controlled.

3.3 Multiplayer Collaboration Mode

Mechanism: each player controls their own character, third-person narration, supporting online and local play.

This is the only platform supporting multiplayer collaboration among the 8 platforms in this group (neither NovelAI nor Sudowrite has a collaboration feature).

The L3 implication of multiplayer mode: narrative control shifts from "one person" to "taking turns advancing" — each player enters an action on their turn, and the model narrates the result uniformly. This is a form of distributed orchestration, the exact opposite of the "orchestrator-worker centralized orchestration" in Claude workflows.

3.4 Sister Product Voyage

Composition of the World Engine: a deterministic mechanics layer (separated from AI narration) tracks health, inventory, currency, geography, relationships, and long-term consequences across thousands of turns.

Other mechanics: skill checks, combat with health and tactical positioning, persistent NPC relationships, and creator world-building tools.

4. Platform Architecture

图 4-1｜AI Dungeon 平台架构：叙事管线与 Voyage 双层机制

数据来源：基于本文分析绘制的示意图。

4.1 Overall Architecture

[玩家输入行动]
      |
      v
[Story Cards / Memory Banks 条件检索]
      |
      +--> [命中卡片] --+
      |                 |
      +--> [近期上下文] -+--> [拼装 Prompt] --> [选定风格模型] --> [叙述输出]
      |                 |
      +--> [世界设定] ---+
                        |
                        v
              [Memories 更新] + [玩家可选分支/重试]

Voyage version adds one layer:

[玩家输入行动]
      |
      v
[World Engine 确定性机制层]  <-- 独立于 AI 叙述
      |  追踪：生命值 / 物品栏 / 货币 / 地理 / 关系 / 长期后果
      v
[机制结果：技能检定、战斗回合、NPC 关系变化]
      |
      v
[AI 叙述层：把机制结果转成文本]

The two-layer architecture is Voyage's most important design: it hands "numbers and rules" to deterministic code, and "description and dialogue" to the model. This is the same principle as "script-as-tool" in Claude workflows — deterministic tasks are not handed to the model.

4.2 Narrative Loop

AI Dungeon's loop is driven by free input — there are no preset options, and players can say anything. This contrasts with MidReal's "a decision point every 300 words": the former is continuous freedom, the latter is beat-guided prompting.

4.3 Voyage's Deterministic Mechanics Layer

The "across thousands of turns" spec is worth noting: it converts the "long-range state" discussed in this group from a unit of "hundreds of thousands of characters" to "thousands of turns", but the essence of the problem is exactly the same — how to avoid losing state in an extremely long interaction sequence.

World Engine offers a different answer from Claude Book:

World Engine is more thorough: Claude Book's state is still "text for the model to read", whereas World Engine's state is "data held by code". The latter cannot be misread or rewritten by the model, because it never passes through the model.

The cost is generality — World Engine can only track the six predefined categories of values and cannot handle unstructured states such as "the protagonist's complex emotional change toward the master". This is a typical determinism-vs-expressiveness tradeoff.

5. Harness Design

5.1 L1 Context Engineering Layer

AI Dungeon's L1 is the typical conditional-retrieval type (Paradigm B): it does not stuff the entire history into context, but maintains an entry library and draws on it by trigger conditions.

The 4,000–32,000 token context window is mid-level within this group:

Shadow Tiers offer higher context — indicating that Latitude treats context capacity as a premium paid point, the same commercial logic as NovelAI's Lorebook capacity tiers.

5.2 L2 Tools and Execution Layer

"Fine-tuning models by play style" is AI Dungeon's most distinctive contribution at the L2 layer. Other platforms offer "Model A / Model B" (performance or cost differences); AI Dungeon offers "a brutal AI / a chaotic AI / a gentle AI" — turning model selection into game-experience selection.

In the Harness context, this design means: the model itself is a tool, and model routing is tool routing.

5.3 L3 Orchestration and Control Layer

AI Dungeon's L3 is "a loop" rather than "a pipeline" — there is no stage division, only a single repeating "input → generate → update" loop. This suits interactive experiences, but not long-form creation.

Voyage introduced turn-based structure (skill checks, combat turns), a step toward "orchestration", but it is still a loop rather than a pipeline.

5.4 L4 Memory and State Layer

AI Dungeon's L4 is "session-based persistence rather than database-backed" — this is the core of its known shortcomings:

Memory degrades in long sessions; players report drift in character consistency, accents, and item details.

Voyage is a direct fix for this shortcoming. World Engine moves the six categories of state from "model memory" into "code data", realizing a fourth approach beyond this group's three long-range paradigms:

Paradigm D applies when state can be numerically represented (HP, gold, item counts, relationship values). For non-numerical states such as "the protagonist's inner change" or "the emotional weight of a foreshadowing", one must fall back to A/B/C.

5.5 L5 Evaluation and Observation Layer

L5 is the weakest layer on this platform. The only "evaluation" comes from complaints in the player community — memory drift, the $49.99/month Mythic tier controversy, and bug reports (infinite loading, text duplication, editing problems).

For comparison: Sudowrite has AI Beta Reading ×3, and Claude workflows have Perplexity Gate + parallel reviewers. AI Dungeon has no corresponding mechanism at all.

This is not a design oversight but a consequence of positioning — the quality of interactive experiences is judged by players in real time; there is neither the need nor the time for automatic evaluation.

5.6 L6 Governance and Safety Layer

The content filter is the core tension of this platform's L6: on one hand it must filter sensitive content (historically the Steam delisting incident), and on the other hand the filtering interrupts dramatic and violent scenes, breaking narrative immersion. This is a governance problem unique to interactive fiction — the granularity of moderation is "a single sentence", while narrative meaning comes from "context".

Compared to NovelAI's "lax content policy", AI Dungeon chose strict filtering; the two user bases have diverged accordingly — creators of restricted subject matter go to NovelAI, while mainstream players stay with AI Dungeon.

5.7 Summary of the Six Layers

Echo of the core thesis: AI Dungeon is an edge case of this group's core thesis. In pure interactive fiction scenarios, the pressure on L4 is greatly relieved by "limited session length" — a session lasts a few hours, and 25–800 Memories suffice. But when Latitude wanted to upgrade the product from a "narrative engine" to a "full emergent RPG" (spanning thousands of turns), L4 immediately became the bottleneck, requiring a five-year investment to develop World Engine.

This conversely proves the core thesis: long-range capability is determined by L4, and the engineering cost of L4 far exceeds expectations — Latitude spent five years.

6. Practical Cases

6.1 Case 1: Voyage's World Engine and Emergent NPCs

Background: AI Dungeon's core shortcoming is that it is "not an RPG in a mechanical sense" — no dice, no character sheet, no skill checks; outcomes are decided by AI rather than by mechanics. The player community has long complained about memory drift.

Solution: Latitude spent five years developing World Engine, using a deterministic mechanics layer (separated from AI narration) to track health, inventory, currency, geography, relationships, and long-term consequences across thousands of turns; invite-only testing opened on iOS and Android on April 21, 2026.

Results (Beta data):

Engineering interpretation:

1. "160,000+ unique NPCs" is direct evidence of emergence — these NPCs were not written by designers, but produced by the interaction of system rules and player behavior.
2. "An average of 3,000 decisions" proves that long-range interaction genuinely occurred — this is exactly the scope of scenarios World Engine is meant to solve.
3. Shared subscription is an important migration-strategy design — letting existing paid users migrate at zero cost, avoiding a cold start failure for a new product.

Significance for the core thesis: Latitude spent five years and a separate product to shore up L4, providing this group's most weighty cost evidence for the core thesis — making L4 solid is not adding a feature, but an engineering effort measured in years.

6.2 Case 2: Memories Tiers and Memory Drift

Background: The community has long reported that "memory degrades in long sessions; players report drift in character consistency, accents, and item details".

Tier design:

Diagnosis:

Key judgment: The tiering of Memories from 25 to 800 shows that Latitude understands "memory capacity" is a core paid point. But capacity cannot solve drift — even with 800 Memories, if state is stored in natural language and passed through the model's context, it will still degrade turn by turn. The real solution is Voyage's deterministic mechanics layer: turning state from "model memory" into "code data".

6.3 Case 3: Migration from Narrative Engine to Full RPG

Background: AI Dungeon's success came from "model as narrative engine", but its limitations also come from this — "outcomes are decided by AI rather than by mechanics".

Migration path:

Current limitations:

• iOS / Android invite-only;
• No Web access;
• No open registration.

Pricing comparison (AI Dungeon vs Voyage):

Note the Mythic tier credit difference: AI Dungeon 2,750, Voyage 3,750. The two pricing schemes are basically aligned with slight discrepancies, presented side by side at the time of writing.

6.4 Counter-example: Credits Economy and Missing Mechanics

Two categories of problems reported by the community:

The shared-credit problem deserves separate analysis: text and images share the same quota pool, forcing players to weigh "should I generate an image this time?" before every interaction. This contrasts with Sudowrite's "Chat-only mode does not consume credits" — Sudowrite makes the "does not consume" judgment explicit, while AI Dungeon leaves users to weigh costs in opacity.

Engineering lesson: the design goal of cost guardrails is to guide behavior, not to create anxiety. Good cost design should make users clear about "what is free, what is charged", rather than making them compute the cost of every operation.

7. Summary

7.1 Strengths

1. The earliest interactive fiction platform in this group (2019), originator of the product form, with 8 million+ registered players.
2. Multi-style fine-tuned models are this group's only "model personalization" design: Muse-12B (relationship-oriented), Wayfarer (combat-heavy), Madness (chaotic), Hearthfire-24B (daily-life atmosphere).
3. Story Cards and NovelAI Lorebook jointly validate the viability of the "keyword-triggered entry library" paradigm.
4. The only platform in this group supporting multiplayer collaboration (online and local; each controls their own character, third-person narration).
5. Voyage's World Engine introduces a fourth long-range state paradigm — the deterministic mechanics layer, where state is held by code and the model only reads results.
6. The world publishing mechanism lets creators publish worlds and scenes for others to explore, forming a nascent UGC ecosystem.
7. The free-tier model is publicly released on Hugging Face, a rare level of transparency.
8. Voyage and AI Dungeon share a subscription, letting existing paid users migrate at zero cost.

7.2 Limitations and Known Shortcomings

1. Not an RPG in a mechanical sense: no dice, no character sheet, no skill checks; outcomes decided by AI rather than mechanics (only added by Voyage).
2. Persistence is session-based rather than database-backed: memory degrades in long sessions; players report drift in character consistency, accents, and item details.
3. The credits economy causes anxiety: text and images share credits; the economy model is hard to grasp.
4. The $49.99/month Mythic tier is controversial in the community; the Ultimate tier is $99.99/month.
5. Bug reports are common: infinite loading, text duplication, editing problems.
6. L5 is almost entirely blank: no automatic quality evaluation, no consistency checking.
7. The content filter interrupts dramatic and violent scenes, breaking narrative immersion (main community complaint).
8. History of content moderation disputes and a Steam delisting incident.
9. Voyage's current limitations: iOS/Android invite-only, no Web access, no open registration.
10. No API / plugins / scripts, weak openness.

7.3 Applicability Boundaries

7.4 Selection Recommendations

• Best-fit scenario: You want to play a story rather than write a work, and want different styles of "AI host". AI Dungeon's multi-style models are the only ones in this group that can satisfy this need.
• Multiplayer scenario: The only platform in this group supporting multiplayer collaboration; irreplaceable.
• Special note on Voyage: If you care about the "deterministic mechanics + emergent narrative" direction, Voyage is worth following, but it is currently invite-only, mobile-only, no Web. And Voyage shares a subscription with AI Dungeon, so existing paid users need no additional payment.
• Not recommended: If your goal is to produce a publishable novel, AI Dungeon is not the right tool — it has no chapter pipeline, no consistency checking, and memory degrades in long sessions. Consider 02-sudowrite.md or 05-claude-fiction.md.
• Tier recommendation: Memories is the only paid point directly tied to L4. If you value long-session coherence, Journey (100) is the lowest usable tier; the free tier's 25 Memories are basically insufficient in long sessions.

Information Gap Declaration

1. Direct verification of the official AI Dungeon pricing page: Forward Future notes it was verified on 2026-08-30, but no original text from the official latitude.io page was seen; marked [To be verified].
2. Voyage's Memories tiers: Whether Voyage uses the same Memories tiers as AI Dungeon was not found; marked [To be filled].
3. Voyage's multiplayer mode: No documentation found; marked [To be filled].
4. Voyage's timeline and versioned-snapshot mechanism: No documentation found; marked [To be filled].
5. Specific specs of Shadow Tiers: The source only states "offers higher context and access to stronger models", with no specific numbers; marked [To be filled].
6. Developer API: No public information found; marked [To be filled].
7. AIGC labeling: No description of a related AI Dungeon implementation found.
8. Inconsistent image-credit tier names: The source lists Adventurer 480, Champion 760, Legend 1,650, Mythic 2,750, which do not fully correspond to the subscription tier names (Wanderer/Journey/Legend/Mythic/Ultimate); marked [To be verified].
9. Specific timing and details of the Steam delisting incident: The source only mentions "a history of content moderation disputes and a Steam delisting incident", with no details; marked [To be filled].
10. Mythic tier credit difference between Voyage and AI Dungeon (2,750 vs 3,750): presented side by side without explanation (source did not state a reason).

8. References

1. Roleforge《Best AI Dungeon Master Tools 2026》 — Roleforge. https://roleforge.ai/blog/best-ai-game-master-tools-compared
2. AI Cloud Base《AI Dungeon》 — AI Cloud Base. https://aicloudbase.com/tool/ai-dungeon
3. Challenging Voice《Latitude.io Review》 — Challenging Voice. https://www.challengingvoice.com/?p=9108/
4. Forward Future《AI Dungeon》 — Forward Future (verified 2026-08-30). https://forwardfuture.com/tools/details/ai-dungeon
5. BuildFastWithAI《NovelAI Review 2026》 — BuildFastWithAI. https://buildfastwithai.com/ai-tools/novelai
6. AI Text Tools《NovelAI: Features, Pricing & Reviews 2026》 — AI Text Tools, 2026. https://aitexttools.net/tools/novelai
7. Apple App Store《MidReal - AI Stories》 — MidReal Inc. https://apps.apple.com/mr/app/midreal-ai-stories/id6566178525
8. HackerNoon《Claude Book: A Multi-Agent Framework for Writing Novels with Claude Code》 — HackerNoon. https://hackernoon.com/claude-book-a-multi-agent-framework-for-writing-novels-with-claude-code
9. Blog《如何用 Agent 写一本小说（其二）》 — tulancn. https://tulancn.github.io/%E5%AD%A6%E4%B9%A0/study/%E5%A6%82%E4%BD%95%E7%94%A8Agent%E5%86%99%E4%B8%80%E6%9C%AC%E5%B0%8F%E8%AF%B4%EF%BC%88%E5%85%B6%E4%BA%8C%EF%BC%89
10. GitHub — forsonny/Claude-Code-Novel-Writer（Multi-Agent Novel Writer v4.1） — 2026. https://github.com/forsonny/Claude-Code-Novel-Writer
11. Test report《用 Claude Code 写小说，意外发现了组队开挂模式》 — 2026. https://m.aitntnews.com/newDetail.html?newId=20196
12. Sudowrite official blog《How Sudowrite Works: Story Bible, Muse, and the Tools》 — Sudowrite. https://sudowrite.com/blog/how-sudowrite-works