NovelAI Platform Research

1. Introduction

1.1 Developer and Release History

NovelAI is the platform in the AI fiction space that first established "memory" as a standalone product concept. When it launched in 2021, the mainstream approach was still "stuff all prior text into the prompt," yet NovelAI already offered a two-tier structure of Lorebook (a persistent worldbuilding entry library) and Memory (near-term event tracking). This design predated the industry's widespread adoption of structured summaries by roughly four years.

1.2 Product Positioning and Target Users

NovelAI positions itself as a subscription-based creative platform, offering both a text generation engine and an anime-style image generator. Its target user base is highly vertical:

• Tabletop RPG players (Text Adventure module)
• Fan-fiction authors (who need to maintain existing worldbuilding and character relationships long-term)
• Visual novel creators (text + image synergy)

What these three user groups share is: sensitivity to "content freedom" and "privacy" far outweighs sensitivity to "publishing compliance." This directly explains NovelAI's choices at the L6 governance layer — it invests resources in encryption and non-training rather than in content moderation.

1.3 In-House Model Lineage

Xialong's release is a notable signal: a company marketing itself on "in-house models" chose to base its newest flagship model on domain fine-tuning of a third-party open-source base (GLM-4.6). This shows that by 2026, the capability gains of general-purpose bases have made "building small models from scratch" poor value for money in creative writing — data (the proprietary creative dataset) is worth more than parameters.

1.4 Pricing and Tiers

Notes:

• Annual billing saves roughly 20%.
• Anlas is a dedicated quota unit for image generation, billed separately from text generation.
• Tablet's and Scroll's Anlas quotas are recorded as 1,000 and 10,000 respectively in different sources — a conflict, marked [To be verified].
• The free tier (Paper) only offers a limited number of uses and does not constitute a genuinely usable free layer.

1.5 Openness and Access

This closed openness is NovelAI's most distinctive business characteristic: it does not seek to become a "platform" but rather an end-to-end closed creative terminal. This is both the technical precondition of its privacy promise (no API means data never leaves the client logic) and the ceiling of its integrability.

2. Glossary

3. Feature Overview

3.1 Core Feature List

3.2 How Lorebook Works

Lorebook is NovelAI's most recognizable design and the industrialized prototype of what this group categorizes as "Long-Run State Management Pattern B · Keyword-Triggered Entry Library."

How it works:

1. Entry definition: The author creates an entry for each character, place, setting, and world rule; each entry contains several trigger words and a body of text.
2. Trigger detection: During generation, the system scans the current context window to see whether any entry's trigger words appear.
3. Conditional injection: Only matched entries are injected into the context, constrained by that tier's memory budget.
4. Budget trimming: If the total length of matched entries exceeds the budget, they are truncated by priority or order.

Capacity (source: BuildFastWithAI):

Engineering implication: Lorebook decouples L4 (persistent storage) from L1 (context assembly) — settings reside in storage and enter context on demand. Compared with "stuffing all prior text in," context usage drops from O(full text) to O(matched entries); compared with "rolling summary," entry content is hand-refined and not compressed by the model, introducing no summary bias.

The cost: trigger hit rate becomes a new failure point. Untriggered settings are equivalent to non-existent — if the term "Qingyun Sect" does not appear in chapter 40, its related settings never enter context, and the model may invent something like a "Qingyun Gate" out of thin air.

3.3 Memory and Author's Note

The division of labor is a clear three-layer structure: Lorebook manages "what the world is," Memory manages "what just happened," and Author's Note manages "how to write now."

Author's Note's implementation approach is noteworthy: it is a meta-instruction inserted at a specific position in the context, not a dialogue message. In Harness terminology, this belongs to L1's "prompt assembly priority" — placing style instructions near the generation point so their weight exceeds that of the body text.

3.4 Sampling and Bias Control

NovelAI exposes these parameters directly to end users — a design for "prompt engineers" rather than "ordinary authors." Multiple reviews describing its UI as "hard to learn" are directly related to this.

Notably: repetition penalty is NovelAI's only built-in anti-repetition measure, yet known shortcomings show it cannot fundamentally fix plot looping in long sessions (see Section 6.4).

3.5 Text Adventure and Branching

Text Adventure Module extends NovelAI from a "writing tool" into a "narrative engine":

• Advances in second person; the player inputs actions and the model writes the outcome;
• Integrates with Lorebook — when the player mentions a location, the corresponding entry is triggered;
• Story branching and parallel timeline exploration is non-destructive; you can return to any node and retry.

This module's engineering value is that it hands control of the L3 orchestration layer to the human player, forming a "human supplies intent, model supplies description" loop. This pattern was later amplified in AI Dungeon and MidReal.

4. Platform Architecture

图 4-1｜NovelAI 平台六层架构：从接入层到治理层

数据来源：基于本文分析绘制的示意图。

4.1 Layered Architecture

4.2 Context Assembly Data Flow

[用户开始生成]
      |
      v
[扫描上下文窗口内的触发词]
      |
      +---> [命中的 Lorebook 条目] --+
      |                              |
      +---> [Memory: 近期事件] ------+
      |                              |
      +---> [Author's Note] ---------+--> [按记忆预算裁剪] --> [拼装 Prompt]
      |                              |
      +---> [正文窗口] --------------+
                                     |
                                     v
                            [采样参数: temperature / top-p / rep penalty]
                                     |
                                     v
                            [模型生成] --> [Biases 后处理倾向]
                                     |
                                     v
                            [Memory 更新] + [加密存储]

The key feature of this data flow is: the context is not "the most recent N tokens," but a heterogeneous assembly of "triggered entries + recent events + meta-instructions + the most recent N tokens." This is precisely the standard form of L1 context engineering in the fiction scenario.

4.3 Storage and Encryption Design

NovelAI's privacy promise consists of four things (source: AI Text Tools / Listicler):

1. XSalsa20 client-side encryption — stories are encrypted at rest, and even employees cannot read user content;
2. No public gallery — there is no path for a work to be publicly browsed;
3. No training on user text by default;
4. No server logs.

Together, this combination forms its core moat in the AI fiction space: for fan-fiction and restricted-theme creators, "the platform can't see it" matters more than "the platform writes better."

Note that client-side encryption is also an engineering constraint — it means the server cannot index, search, or cross-analyze content, so NovelAI could never offer capabilities like "whole-book foreshadowing search" that require the server to read full text. This is a clear trade-off between privacy and functionality.

4.4 Conflicting Context Window Figures

Public sources describe NovelAI's context window very differently, and the official documentation gives no unified figure, so the figures must be presented side by side:

Possible explanations for the discrepancies (inferred, not official):

• Different sources may be measuring different subsystems — story memory, Lorebook capacity, and the model context window are three different things;
• Different sources may be measuring different models — Kayra and Xialong have different windows;
• Different sources may be measuring at different points in time, and the window may have jumped after Xialong (2026-03) was released.

Conclusion: any statement citing a single number is unreliable. This group uniformly marks these as [To be verified].

5. Harness Design

5.1 L1 Context Engineering Layer

NovelAI's L1 is a textbook "heterogeneous assembly under a budget constraint," with four components each having a clear responsibility:

Compared with the other platforms in this group, NovelAI's L1 has two characteristics:

1. It does not rely on rolling summaries. It has no mechanism for compressing prior text into summaries; instead it relies on Lorebook's "hand-refined entries" to carry long-range information. The upside is no compression bias; the downside is that the entire maintenance burden falls on the author — the author must write out every possibly-relevant setting as an entry in advance.
2. Budget tiering is bound to the subscription plan. Lorebook capacity ranges from 200 tokens on Tablet to 2,048 tokens on Opus — a 10x difference. This is a design that puts L1's core capability behind a paywall, and it means lower-tier users effectively cannot maintain complex worldbuilding.

5.2 L2 Tooling and Execution Layer

L2's shortcoming is very clear: no API, no plugin ecosystem. This contrasts sharply with Sudowrite's 1,000+ community plugins. NovelAI's toolset is "sufficient but not open" — all capabilities are built into the client, so authors cannot integrate external scripts (such as a name generator, word-count tool, or foreshadowing ledger) into their workflow.

5.3 L3 Orchestration and Control Layer

NovelAI's orchestration capability ranks in the middle among the 8 platforms:

NovelAI's L3 is "human-driven" rather than "pipeline-driven": there is no automatic plan → draft → review loop; at every step the author decides what to do next themselves. This gives the author maximum control, but it also means the author must act as the orchestrator — which explains why multiple reviews mention a "steep prompt-engineering learning curve."

5.4 L4 Memory and State Layer

This is the strongest layer of NovelAI and the one that best embodies this group's core thesis.

Key judgment: NovelAI's L4 is a state layer of "strong settings, weak temporality." It solves "what the world is like" extremely well (Lorebook), but barely addresses "what has happened so far" (no timeline, no State snapshots, no foreshadowing ledger).

This is exactly why prose repetition and plot looping still occur in long sessions: the model knows "what Qingyun Sect is like," but does not know "the protagonist already visited Qingyun Sect three chapters ago" — the latter is temporal state, which falls outside Lorebook's coverage.

Comparing with Claude Book's State snapshot design reveals the gap: what NovelAI lacks is precisely the step of "extracting and archiving changes in character location, possessions, knowledge, and relationships after each chapter."

5.5 L5 Evaluation and Observation Layer

L5 is NovelAI's weakest layer. It offers only pre-generation tendency controls (Biases, sampling parameters) with no post-generation automatic evaluation. Judging whether something is "well-written" falls entirely on the author.

The cost of this gap is amplified in long-form scenarios: authors cannot automatically discover that "the character settings have drifted by chapter 40" or that "this passage has slid noticeably toward the AI Slope" — they can only catch it by re-reading.

5.6 L6 Governance and Security Layer

NovelAI's L6 is privacy-first rather than compliance-first, which is the most distinctive among the 8 platforms in this group:

It should be stated explicitly: NovelAI's lenient content policy and Chinese compliance frameworks such as the "Measures for the Labeling of AI-Generated Synthetic Content" are not in the same applicable context (it mainly serves the English market). But for Chinese users, if a work is to be published on a Chinese platform, NovelAI provides no AIGC labeling capability at all, so the full compliance obligation falls on the author.

5.7 Six-Layer Capability Summary

Echo of the core thesis: NovelAI used Lorebook to make L4's "setting dimension" solid, but the "temporal dimension" remains a blank. This precisely confirms this group's thesis — the ceiling of long-form capability is determined by L4, and L4 must be complete in both "settings + temporality." NovelAI only did half, so it excels at medium-short fiction and interactive narrative but still has a structural shortfall in ultra-long-form work.

6. Practical Cases

Note: NovelAI has not published any official customer case study with quantitative metrics. The three cases below are scenario extrapolations based on publicly available feature combinations, used to illustrate its capability boundaries; they are not vendor-disclosed case studies. Section 6.4 covers known issues that repeatedly appear in public reviews.

6.1 Case 1: Maintaining Long-Form Fanfiction Worldbuilding

Background: The core constraint of fan-fiction is that you must "stay within the original work's worldbuilding" — the original's geography, factions, character relationships, and ability systems are all fixed, and any mistake is immediately pointed out by readers. The L4 pressure of such creative work concentrates on the setting dimension rather than the temporal dimension.

Solution (extrapolated from publicly available features):

1. Create a Lorebook entry for each major location, faction, ability, and key character in the original work, using their common names and aliases as trigger words;
2. Use Author's Note to fix the narrative point of view and writing style (e.g., "second person, restrained and terse");
3. Use Biases to suppress the author's verbal tics and reinforce the original work's distinctive imagery words;
4. On the Opus tier, use the 2,048-token Lorebook budget to accommodate a fairly complete setting collection.

Results and boundaries: This is an optimal use case on NovelAI's capability curve — dense settings, simple temporality, and controllable entries. Lorebook's "on-demand injection" here both saves context and ensures accuracy. But if a fanfic serializes past several hundred thousand characters and the number of side characters grows significantly, the Lorebook budget quickly peaks, and the lack of temporal state causes errors such as "when did a certain person know a certain thing."

6.2 Case 2: Interactive Adventure Branching

Background: Text Adventure Module targets the loop of "player inputs an action — model continues the result," which requires the model to respond to new elements introduced by the player in real time.

Solution (extrapolated from publicly available features):

1. Preload Lorebook entries for world rules and key NPCs;
2. The player inputs freely; when a trigger word hits, the corresponding entry is injected immediately;
3. Use non-destructive branching to save parallel timelines at key decision points, rolling back and retrying after failures;
4. Lower temperature to improve world consistency; raise repetition penalty to suppress repetition in scene descriptions.

Results and boundaries: The non-destructive branching design is this scenario's key advantage. But known shortcomings show that under long sessions the model still loops back to similar scenes; repetition penalty can only ease wording repetition, not stop plot repetition — because "plot repetition" is a symptom of missing temporal state, not a sampling-parameter problem.

6.3 Case 3: Fine-Tuning Writing Style with Custom Modules

Background: Custom Module Training (available on the Opus tier) lets users train a module on their own text.

Solution (extrapolated from publicly available features):

1. Collect sample documents of the target style (e.g., an author's existing works);
2. Train a custom module so the generation distribution converges toward that style;
3. Combine it with Lorebook — the module handles style, Lorebook handles settings;
4. Use Author's Note for final fine-tuning within a session.

Results and boundaries: This is a distinctive capability that sets NovelAI apart from other subscription tools at the L2 layer. It is essentially the individualized version of the same idea behind Xialong's "fine-tuning on the proprietary creative dataset over GLM-4.6" — domain-data fine-tuning is more stable than prompt engineering. The risks: if the training data contains copyrighted works, ownership disputes may arise; and the degree of similarity between the trained module's output and the original style is hard to quantify.

6.4 Counterexample: Prose Repetition and Plot Looping in Long Sessions

A problem that repeatedly appears in public reviews: even with Lorebook and Memory in place, prose repetition and plot looping still occur under long sessions.

Diagnosis (analysis based on this group's six-layer framework):

The value of this counterexample is that it proves a single powerful setting library (Lorebook) cannot support long-form work — temporal state management must also be present. This is the most direct negative evidence for this group's core thesis.

7. Summary

7.1 Strengths

1. Lorebook is the earliest and most mature keyword-triggered context-injection system of its kind, decoupling the L4 persistence layer from the L1 assembly layer; this design was later inherited by AI Dungeon (Story Cards).
2. The privacy promise is complete and technically verifiable: XSalsa20 client-side encryption + no public gallery + no training by default + no server logs form a moat in the restricted-theme creative domain.
3. Text + image + TTS in one, providing a closed loop for visual novel creators.
4. A clear in-house model + domain fine-tuning path: from Kayra (13B) to Xialong (GLM-4.6 fine-tune) shows it builds barriers on the data side rather than the parameter side.
5. Non-destructive branching, supporting parallel-timeline experimentation.

7.2 Limitations and Known Shortcomings

1. No API, no mobile app, no collaboration features — the worst openness; it cannot be integrated into larger workflows.
2. UI is hard to learn, with a steep prompt-engineering learning curve (a consensus across multiple reviews).
3. Prose repetition and plot looping under long sessions, even with the memory system; the root cause is that L4 lacks temporal state.
4. No timeline, no State snapshots, no foreshadowing ledger — harmless for medium-short works, a hard shortfall for long-form ones.
5. L5 is nearly blank: no automatic evaluation capabilities at all.
6. Image quality lags Midjourney (except in the anime sub-niche).
7. No truly free tier: the Paper tier only offers 50 generations.
8. Lorebook capacity is tightly tied to the subscription tier; the Tablet tier's 200 tokens effectively cannot maintain a complex worldbuilding.

7.3 Applicability Boundaries

7.4 Selection Recommendations

• Criteria for choosing NovelAI: your core need is to "keep generating within a self-consistent world" and "not want the platform to see the content." If both hold, NovelAI is the optimal choice in this group.
• Criteria for not choosing it: if your project exceeds 300,000 characters, needs foreshadowing recovered across chapters, or needs team collaboration or external tool integration, NovelAI's structural shortfalls grow with length. In that case, prioritize the self-built workflow described in 05-claude-fiction.md.
• Tier recommendation: if you decide to use it, the Lorebook budget is the only unavoidable expense — Tablet (200 tokens) is insufficient for a complex worldbuilding; Scroll or above is recommended. Xialong is available only on the Opus tier, so if the latest model matters to you, you must choose the $25/month tier.

Information Gap Statement

1. Context window: five different figures exist — 2,048 / 6,144 / 8,192 / 28,672 / 128k tokens — and the official documentation does not unify them, so they are presented side by side and marked [To be verified]. The differences may stem from different tiers, different models (Kayra / Xialong), and different subsystems (story memory / Lorebook / model context).
2. Tablet and Scroll Anlas quotas: different sources record 1,000 and 10,000/month respectively — a conflict, marked [To be verified].
3. Clio model's parameter scale and base information: no public description found, marked [To be filled].
4. Copyright claims: no explicit statement by NovelAI about rights over generated content found, marked [To be filled].
5. AIGC labeling compliance: no implementation description relevant to the "Measures for the Labeling of AI-Generated Synthetic Content" or similar regulations found.
6. Official customer case studies: no official cases with quantitative metrics found; Cases 1–3 in Section 6 are scenario extrapolations based on public feature combinations, clearly labeled as such.
7. Monthly traffic and market share: ToolNest records 4M monthly visits — a single source, not cited as a fact in the body text.

8. References

1. AI Text Tools, "NovelAI: Features, Pricing & Reviews 2026" — AI Text Tools, 2026. https://aitexttools.net/tools/novelai
2. RECATOOLS, "NovelAI Review" — RECATOOLS. https://recatools.com/ai-directory/novelai/
3. Listicler, "NovelAI" — Listicler. https://listicler.com/tools/novelai
4. BuildFastWithAI, "NovelAI Review 2026" — BuildFastWithAI. https://buildfastwithai.com/ai-tools/novelai
5. ToolNest, "NovelAI" — ToolNest, 2026. https://toolnest.ai/project/novelai
6. Fanqie Novel official announcement, "Notice on the Launch of the AI Writing Tool Feature" — Fanqie Novel, 2024. https://fanqienovel.com/writer/zone/article/7327136545129906238
7. Haike Finance, "Fanqie Novel's AI Problem" (Sina Finance) — Haike Finance, 2025. https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml
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. Roleforge, "Best AI Dungeon Master Tools 2026" — Roleforge. https://roleforge.ai/blog/best-ai-game-master-tools-compared
10. Sudowrite official blog, "How Sudowrite Works: Story Bible, Muse, and the Tools" — Sudowrite. https://sudowrite.com/blog/how-sudowrite-works
11. Toutiao, "Novel Writing Technique Notes: From Beginner to Advanced" — Toutiao. https://m.toutiao.com/article/7666250379276927531
12. Toutiao, "A Complete Method for Planning a Full-Length Novel Outline" — Toutiao. https://m.toutiao.com/article/7638890425910968851
13. Meipian, "How Beginners Start Writing Web Novels" — Meipian. https://www.meipian.cn/5gn73n9w
14. Hands-on article, "Writing a Novel with Claude Code and Accidentally Discovering a Team Synergy Mode" — 2026. https://m.aitntnews.com/newDetail.html?newId=20196