MidReal Platform Research (Interactive Narrative)

1. Introduction

1.1 Developer and Basic Information

MidReal is the younger of this group's two interactive-narrative platforms (the other is AI Dungeon from 2019, see 06-ai-dungeon.md). Its form is not a "writing tool" but an interactive story generator built for mobile: the user gives a simple concept, the platform expands it into a full story in about 10 seconds, and hands narrative control back to the user at fixed beats. From a Harness perspective, it compresses the form of "reader-as-player, narrative generated in interaction" into a 41.1 MB mobile app.

1.2 Product Positioning and Target Users

MidReal is positioned as an interactive-narrative / AI interactive-fiction platform, with a core experience loop of:

1. Input a concept (covering genres such as horror, fantasy, sci-fi, romance, and more);
2. AI expands the story, pausing roughly every 300 words and returning control;
3. The user decides the narrative direction, or lets AI rewrite unsatisfactory parts;
4. The story can continue indefinitely in length; along the way images can be generated and conversations can be had with the protagonist.

Its target users differ fundamentally from this group's "author-type" platforms (Sudowrite, 灵蟹创作): MidReal targets "player-type readers" — people who don't aim to produce submittable works but instead consume "a narrative experience driven by their own decisions." This shapes its Harness design focus away from L5 (quality evaluation) and toward L3 (interaction orchestration) and L4 (state consistency).

1.3 Development Team Background: A Single-Source Problem

The third-party toolbox LBB.AI claims MidReal was co-developed by MIT, NYU, Cambridge, and Princeton, and labels the region as Australia (collected 2025-05-30).

Handling principle: this claim appears only in that single aggregator source; neither MidReal's official channels nor the Apple App Store page confirm it, making it a typical "aggregator-site marketing description." This report consistently labels it single source. Until there is official disclosure, any statement citing this background cannot be treated as fact. The only developer information that can be confirmed is the two entity names, MINDAI PTE. LTD. and MidReal Inc., from the App Store page.

1.4 Pricing: A Conflicting Basis of Two Coexisting Naming Systems

MidReal's public pricing has two mutually incompatible naming systems that appear together on the same App Store page, likely reflecting old and new pricing coexisting. This report presents them side by side without choosing between them, consistently marked [To be verified].

Basis 1: The "Dose" series on the App Store in-app purchase page (relevance: high)

Basis 2: The official site's upgrade-plans as retold by Toolify (relevance: medium)

Basis 3: GETAI.APP's Chinese site (relevance: low): free version ¥0; standard version ¥5.99/month; pro version ¥29.99/month.

Three observations:

1. The price anchors are highly consistent: the $5.99 / $14.99 / $29.99 tiers in all three bases correspond exactly, indicating that the "three-tier pricing" itself is credible — the conflict is only in naming and allowance details;
2. "Dose" and "Pills" share a single drug metaphor, an unusual naming style implying its allowance system is designed around "generation dose," belonging to the same "consumable billing" paradigm as AI Dungeon's Credits and NovelAI's Anlas;
3. Premium/Ultra and Low/Medium/High Dose appear side by side on the App Store, making it impossible to confirm which set is the current price; any citation should note the dual basis.

1.5 Open Form and Access Methods

Unlike NovelAI's "closed for privacy" and Fanqie's "closed for governance," MidReal's open form falls between the two: it has community and sharing mechanisms (reading others' stories, Discord), but its engineering-side openness (API, plugins) lacks verifiable information at the official level digitally, and cannot be confirmed the way Sudowrite's 1,000+ plugin ecosystem can.

2. Glossary

3. Feature Description

3.1 Core Feature List

A key observation on the feature structure: MidReal's feature list contains no author-side concepts such as "outline," "detailed outline," "foreshadowing," "Story Bible," or "Lorebook." All of its features are designed around the "consumer-side experience"—expansion, decision, redirection, illustration, and dialogue. This forms a structural divergence from this group's author-type platforms and directly determines the pressure distribution across Harness's six layers discussed in section 5.

3.2 Control Returned Once Every 300 Words

This is MidReal's most recognizable design. Compare the "human in the loop" beats of this group's other platforms:

The engineering implications of a fixed 300-word beat are threefold:

1. For the user: it removes the decision burden of "should I interfere"—the system evens out the insertion timing for you, suiting mobile's fragmented consumption scenarios;
2. For the state layer: every choice point is a candidate location for a world-state fork. A 300-word interval means a 10,000-word story naturally produces about 30 potential fork points, and state-management pressure stacks linearly with story length;
3. For the context: generation after each choice point must re-assemble "the state line that is still valid across all branch history," which is the direct source of the life-or-death problem discussed in section 5.5.

3.3 Memory Span

The App Store page lists the "Memory Span technology" as the core capability for ensuring long-term consistency and avoiding contradictions. It must be stated honestly:

• The official channels gave only the name, disclosing no implementation details: whether it is a rolling summary, a keyword-triggered entry library, or versioned state snapshots (the three long-range state paradigms A/B/C as grouped by this group, see this group's README section 3.4), the public materials cannot tell;
• Against industry facts: NovelAI uses Lorebook (Paradigm B), Claude Book uses State snapshots (Paradigm C), and Claude Code uses Rolling Summary (Paradigm A)—all three paradigms have auditable public implementations, whereas Memory Span can currently only be cited as a vendor claim;
• Therefore, all descriptions of Memory Span's capabilities in this report are limited to "official claims," and no public quantitative data on its actual effect was found.

3.4 Character Chat and Role Play

MidReal offers two auxiliary ways to "enter the story":

1. Chat with the protagonist: bypass the narrative process and chat directly with a character in the story. In Harness terminology this is a side session independent of the main narrative loop—it raises a classic state-consistency question: does information gained in the chat get written into the world state? If written, does the main narrative acknowledge it? If not, the chat becomes a "consequence-free illusion." Public materials do not state how MidReal handles this, [To be filled];
2. New plot ideas: let AI proactively propose directions, which in effect outsources part of the decision power of "Change Plot Direction" to the model—a "machine proposal + human adjudication" form at the L3 orchestration layer.

4. Platform Architecture

4.1 Layered Architecture

图 4-1｜MidReal 平台分层架构（基于公开功能推演）

数据来源：基于本文分析绘制的示意图。

Since MidReal has not disclosed an architecture document, the table below is an inferred layering based on public feature descriptions, with each row's evidence strength marked:

Compared with NovelAI and AI Dungeon, the amount of architecture information MidReal discloses is the lowest of the three—this group's other two interactive-narrative documents can at least confirm a model lineup (AI Dungeon's Muse-12B / Wayfarer, etc.) or a model lineage (NovelAI's Kayra / Clio / Xialong), whereas MidReal's underlying model is entirely unknown.

4.2 Interaction Loop Data Flow

An inferred loop based on public feature descriptions:

[用户输入一个概念]
        |
        v
[语义展开: 概念 -> 完整故事开场]        <- 宣称约 10 秒
        |
        v
[生成约 300 词叙事片段]
        |
        v
[到达选择点: 系统暂停, 交还控制权]
        |
        +--> [用户选择走向]  --+
        |                      |
        +--> [Change Plot Direction 开辟新分支] -+
        |                      |                 |
        +--> [要求重写本段] ---+                 |
        |                      |                 |
        +--> [让 AI 提议走向] -+                 |
        |                                        |
        v                                        v
[Memory Span 更新长程状态] <---------------------+
        |
        v
[继续生成] --> [循环至无限长度 / 用户结束]

This loop differs fundamentally from AI Dungeon's "player-input-driven" loop: AI Dungeon's pause points are defined by player input behavior (if the player doesn't speak, the story doesn't move), whereas MidReal's pause points are defined by text length (the system stops when it counts 300 words). The former is input-driven orchestration; the latter is length-driven orchestration—the latter is friendlier to the state layer (pause points are certain and predictable) and more rigid about pacing control.

4.3 Mobile Form and Distribution

MidReal uses the App Store as its primary distribution channel (version records and in-app purchases are all based on the App Store), in sharp contrast to this group's other platforms:

The mobile-first form and its "300-word beat" are mutually causal: the reading volume of a single mobile-screen session naturally approaches a few hundred words, and a fixed beat precisely matches the attention unit of mobile scenarios. This is an example of consistency between product form and orchestration design.

5. Harness Design

5.1 Core Challenge: Branch State Explosion

The fundamental difference between interactive narratives and this group's author-type platforms can be summed up in one sentence: the reader of an author-type platform always follows a single line (the one the author wrote), while in interactive narratives every choice point can fork the world state.

Suppose a story has n choice points and an average of 2 options at each; the number of potential world states is O(2^n):

• 10 choice points → up to 1,024 potential state lines;
• 30 choice points (roughly a 10,000-word story) → up to about 1 billion, although in practice only one mainline path and its adjacent branches are actually reached.

This brings three engineering problems that author-type platforms don't have:

1. The state-ownership problem: if the user picks A at the 3rd choice point, is an event that "only happens if B is chosen" still allowed into the context? (Compare the conclusion of Chinese practitioners: the knowledge graph retains only current-progress information to prevent future information from being disclosed early—see 05-claude-fiction.md);
2. The backtracking-consistency problem: if the user backtracks from choice point 7 to choice point 4 and switches to B, all subsequent state must be recomputed on the basis of "following the B line after point 4," and any residual A-line state causes "plot retconning";
3. The context-budget problem: all generated text cannot be stuffed into the context (indefinitely long stories are doomed to not fit), so the fact set of the "currently active branch" must be maintained in external state.

This group's README core assertion (L4 determines the long-form ceiling) is further strengthened in the interactive-narrative scenario into: branch state management is the life-or-death line for interactive-narrative platforms. MidReal responds to this challenge with Memory Span + unlimited-length stories, but the implementation is not disclosed, and this uncertainty must be preserved throughout the layer-by-layer analysis below.

5.2 L1 Context Engineering Layer

Comparison basis: NovelAI's L1 has Lorebook capacity numbers (200–2,048 token), and AI Dungeon has paid-tier window numbers of 4,000–32,000 token, whereas MidReal's L1 has only one term (Memory Span). On L1 transparency, MidReal is the lowest of this group's six platforms—this doesn't necessarily mean its implementation is worse, but it means outsiders cannot audit its consistency capability, leaving users to judge by "whether it actually forgets in use."

Inferred from the interaction loop, MidReal's L1 must at least handle three kinds of assembly—state injection at choice points (current-branch facts), opening setup (genre, viewpoint, style), and user meta-instructions (redirection commands, rewrite commands)—and its budget strategy has no public information, [To be filled].

5.3 L2 Tools and Execution Layer

MidReal's L2 shows a distinctly consumer-grade toolset character: compared with Sudowrite's 1,000+ plugins and Chinese practice's "scripts-as-tools" (naming scripts, word counting, knowledge-graph queries), MidReal has no productivity-oriented authoring tools (no outline tool, no setting management, no export). Every one of its tools serves "letting players enjoy consuming stories more," and the toolset is highly self-consistent with the product positioning.

5.4 L3 Orchestration and Control Layer

L3 is the most solid layer in MidReal's public information:

The combination of the 300-word beat and Change Plot Direction in effect forms a lightweight state machine: every choice point is a state-transition node, "redirection" is a forced jump, and "rewrite" is a same-state retry. It has no complex multi-agent orchestration, but it makes the position where human decisions are embedded in the loop very explicit—this is the most rigid and most predictable human-in-the-loop beat defined across all of this group's platforms.

5.5 L4 Memory and State Layer (Life-or-Death Line)

As described in section 5.1, L4 is the life-or-death line for interactive-narrative platforms. Let me check MidReal's public information item by item against the state-category list in this group's README section 4.4:

Key judgment: MidReal's publicly disclosed L4 information consists only of two terms, "Memory Span" and "unlimited length," and precisely lacks branch state management—the most core engineering component of interactive narratives. There are two possible explanations (both inferences, not facts):

1. The linearization strategy: MidReal doesn't actually maintain a branch tree—after the user "redirects," old branches are discarded or folded into a summary, and the system maintains only the currently active line. This fits the product temperament of the 300-word fixed beat (lightweight, mobile) and has a low implementation cost, but it means true-branch play such as "going back to the 3rd choice point and choosing again" may be unsupported or only limitedly supported;
2. The black-box built-in strategy: branch state management is built into Memory Span but simply not disclosed. If so, its engineering complexity is no less than Claude Book's State snapshot system.

Whichever explanation holds, it should be made clear: at the level of public information, MidReal cannot be confirmed to have production-grade branch state management. This is the largest source of uncertainty when comparing it with AI Dungeon (Story Cards + Memories tiers, and its sister product Voyage's deterministic World Engine state layer already launched).

5.6 L5 Evaluation and Observation Layer

No public results were found. The search report explicitly records "no public search result → [To be filled]" for this item.

Compare this group's four-layer L5 evaluation framework (mechanical-signal layer / consistency layer / style layer / literary-quality layer, see this group's README section 4.5):

• Mechanical-signal layer: no public evidence (no quality-check-style tool information);
• Consistency layer: no public reviewer / validation mechanism, only Memory Span's indirect claimed coverage;
• Style layer: no public AI Slope countermeasure (compare Perplexity Gate);
• Literary-quality layer: no public AI Beta Reading-style mechanism.

Conclusion: MidReal's L5 is completely blank at the level of public information. Users have only two paths to perceive quality: reading it themselves and checking community feedback. This is self-consistent with its consumer-grade positioning (the player is the evaluator), but it means serious creation scenarios that require quality assurance should not choose it.

5.7 L6 Governance and Safety Layer

L6 judgment: MidReal's governance openness is in this group's lowest tier. Copyright ownership is the only right promised to users, and even that promise has only a third-party source; the question of "whether user stories are used to train the model"—which is a core issue for both NovelAI (explicitly does not train) and Fanqie (the 2024 terms dispute)—has no official answer at MidReal. For privacy-sensitive users, this should be treated as a decision barrier.

5.8 Six-Layer Capability Summary

A response to the core assertion: MidReal delivers this group's most rigid human-in-the-loop beat at L3, but at L4 it only delivers a "term-level" promise. Per this group's assertion (L4 determines the long-form ceiling; L4 is the life-or-death line for interactive narratives), MidReal's current public evidence is insufficient to support the conclusion of "production-grade branch state management"—it more resembles a consumer-grade product that leaves the consistency problem to its own black box and the user's tolerance.

6. Case Studies

6.1 Status of Case Evidence

No public quantitative cases were found for MidReal: no vendor-disclosed user data, no third-party measured test reports, no community-verifiable scale data (compare AI Dungeon's 8 million registered players, Voyage Beta's 160,000+ NPCs and nearly 3,000 decisions per capita). Therefore, the three cases in this chapter are all application-scenario inferences based on the combination of publicly available features, used to illustrate capability boundaries—they are not vendor-disclosed cases and constitute no evidence of effectiveness.

6.2 Case 1: Cold-Starting from One Sentence to an Interactive Story (Scenario Inference)

Background: On a commute, a user wants to play a "cyberpunk detective" interactive story, with 10 minutes of available time and no presets.

Approach (inferred from publicly available features):

1. Input the concept "an amnesiac cyberpunk detective searching for his identity in a neon city";
2. Wait about 10 seconds to get a complete story opening;
3. Every 300 words, choose among the directions the system offers, or tap "change direction";
4. A single-beat consumption duration aligns with a subway-commute unit.

Effect and boundaries: this is the optimal scenario on MidReal's capability curve—short consumption units, no state baggage, and moderate decision density. The 10-second cold start + fixed beat are highly friendly to fragmented scenarios. The boundary: once a single session accumulates beyond a few thousand words, the consistency of branch state depends entirely on the undisclosed Memory Span implementation, and public information cannot give a reliability judgment.

6.3 Case 2: Branch Exploration and Plot Redirection (Scenario Inference)

Background: a player is not satisfied with the options the system offers and wants the story to "completely follow a path no one has taken."

Approach (inferred from publicly available features):

1. At a choice point, abandon the conventional option and use Change Plot Direction to force a redirection;
2. When a more drastic twist is needed, let AI proactively provide new plot ideas and then adjudicate;
3. When unsatisfied with a paragraph after redirection, use rewrite.

Effect and boundaries: redirection is the greatest power MidReal gives players, and also the direct trigger of branch state explosion. The clearly definable capability boundaries in the inference: whether the new branch's old-fact isolation after redirection (old-line events no longer appearing) is reliable, and whether one can roll back to the node before redirection, all have no public information [To be filled]. If the platform actually adopts the linearization strategy (see inference 1 in section 5.5), "backtracking after multiple redirections" may be unusable or degraded in experience.

6.4 Case 3: The Memory Test in Long Continuous Reading (Scenario Inference)

Background: a player advances the same story across several consecutive days, accumulating tens of thousands of words, trying to test "whether a supporting character mentioned on day 2 still remembers the agreement with the protagonist on day 10."

Approach (inferred from publicly available features):

1. Keep consuming and, at key plot points, use the chat-with-protagonist side session to confirm information;
2. Observe whether long-range facts (names, agreements, held items) drift.

Effect and boundaries: this scenario directly tests the true caliber of Memory Span, but this report cannot reach a conclusion—no public third-party long-range consistency measurement was found. What can serve as a comparison is industry fact: NovelAI (with Lorebook) still shows plot loops in long sessions; AI Dungeon's memory degrades in long sessions, with players reporting drift in character consistency and item details. That is the industry baseline; if Memory Span genuinely outperforms that baseline, it would be a significant differentiator, but that conclusion currently cannot be supported by public evidence.

7. Summary

7.1 Strengths

1. Most rigid human-in-the-loop beat design: returning control every 300 words is the most clearly defined, most predictable human-in-the-loop mechanism in this group, naturally matching mobile consumption units.
2. Fastest cold-start experience: expands from a one-sentence concept into a full story in about 10 seconds, the strongest in this group for "quick inspiration and short works" scenarios (same basis as this group's README section 7.1 selection matrix).
3. Complete player-power tools: redirection (Change Plot Direction), rewrite, machine proposals, and chat with the protagonist—four controls cover the main needs of "deciding the direction."
4. Unique mobile-first form: the only interactive-narrative platform in this group that treats the iOS App as a first-class citizen; the lightweight 41.1 MB package lowers the barrier to trying it.
5. Product promise of unlimited-length narrative: stories are not limited by length, theoretically supporting very long-line consumption.

7.2 Limitations and Known Shortcomings

1. Lowest transparency in this group: underlying model, context window, state-management implementation, and training policy are all undisclosed; the control group AI Dungeon at least discloses a lineup of five gameplay-fine-tuned models.
2. No public evidence for branch state management: the life-or-death engineering of interactive narratives (branch tree, snapshots, backtracking) has only a term-level promise (Memory Span).
3. L5 is completely blank: no public evaluation mechanism of any kind; quality perception relies entirely on player subjectivity.
4. Confusing pricing basis: two naming systems coexist on the App Store; third-party bases both complement and contradict each other, raising the cost of purchase decisions.
5. No author-side capability: no outline, no setting management, no export—it cannot play the role of a creative productivity tool.
6. Untrustworthy source for key background: "co-developed by MIT/NYU/Cambridge/Princeton" is a single source; the copyright-ownership promise is likewise only a third-party retelling.

7.3 Applicable Scope

7.4 Route Comparison of Interactive Narratives with AI Dungeon

This group's two interactive-narrative platforms represent two different routes, with differences running through all six layers:

Route judgment: AI Dungeon follows a "thick state layer + mechanization evolution" route—over seven years it moved from a pure narrative engine to the World Engine's deterministic state layer, essentially responding to the branch-state-explosion problem (Voyage uses a deterministic mechanism layer to strip HP, inventory, and relationships from AI's uncertain output, tracking across thousands of turns). MidReal follows a "light interaction + black-box consistency" route—not disclosing its state layer, not building a mechanism layer, folding complexity into the product's black box, and using a rigid beat to bound single-state complexity. The former is heavy-engineering evolution; the latter is consumer-grade packaging.

From the perspective of Harness's six-layer model: AI Dungeon's route aligns with the engineering direction of "state externalized and auditable"; MidReal's route abandons external verifiability at three key layers, L1/L4/L5. If interactive narrative is treated as a serious Harness engineering problem, AI Dungeon (and its Voyage evolution) is the more informative research subject; if it is treated as a mobile consumer product, MidReal's beat design is instead the more elegant product answer.

7.5 Selection Recommendations

• Criteria for choosing MidReal: what you want is "to spend 10 minutes on your phone playing a story you decide yourself," you have basic tolerance for consistency and no requirement for transparency. In that case its cold-start speed and beat design are this group's best experience.
• Criteria for not choosing MidReal: you want to write submittable novels (choose Sudowrite or 灵蟹创作), build an engineered long-form project (choose a Claude self-built workflow), play tabletop-style mechanics (wait for Voyage's public beta or use AI Dungeon), or create in a Chinese-compliance environment (choose the Fanqie / Yuewen toolchain).
• If choosing between the two interactive-narrative platforms: value state verifiability and model selection → AI Dungeon; value mobile experience and decision beat → MidReal.
• What to watch: the disclosure of Memory Span's implementation details, the convergence of the dual pricing names, the launch of Chinese support (claimed by end of 2026), and the officialization of copyright terms—these four signals determine whether it is worth the investment to upgrade.

Information Gap Declaration

The following items did not receive citable authoritative results in this search; all have been marked in the body text, with no speculative filling-in:

1. Development team background: "co-developed by MIT, NYU, Cambridge, and Princeton" appears only in the single LBB.AI aggregator source; neither the official channels nor the App Store page confirm it, marked single source.
2. Pricing naming conflict: the App Store simultaneously lists two in-app purchase naming systems, "Premium / Ultra" and "Low / Medium / High Dose," likely old and new versions coexisting; the current price cannot be confirmed, presented side by side and marked [To be verified].
3. L5 evaluation mechanism: the search found no public results at all, marked [To be filled] (same origin as item 13 of this group's README information gaps).
4. Underlying model and context window: no model name, base, or window number disclosed; the control groups NovelAI and AI Dungeon both have public model lineages.
5. Branch state management implementation: Memory Span is only an officially claimed term; branch tree, snapshot, and backtracking mechanisms all have no public evidence; the two explanations in section 5.5 are both inferences and have been explicitly flagged.
6. Public quantitative cases: no official or third-party measured test data with metrics was found; the cases in section 6 are all scenario inferences and have been marked as such.
7. Copyright ownership terms: "generated-content copyright belongs to the user, free for commercial use" appears only in two low-relevance sources, GETAI.APP and 我的 AI 导航, with no official terms page to corroborate it, marked [To be verified].
8. Model training policy: whether user stories are used to train the model has no public statement of any kind.
9. Web, API (docs.midreal.ai/api), and example library (github.com/midreal/examples): appear only per GETAI.APP's retelling, not officially confirmed, marked [To be verified].
10. Chinese support: "expected to launch by end of 2026" appears only in third-party sources, marked [To be verified].
11. State consistency of the side session: whether information from chatting with the protagonist is written into the world state has no public explanation.
12. Direct verification of the official site and official docs: all of this report's high-relevance facts about MidReal come from the Apple App Store page; the original text of the official midreal.ai site could not be verified.

8. References

1. Apple App Store《MidReal - AI Stories》 — MidReal Inc.。https://apps.apple.com/mr/app/midreal-ai-stories/id6566178525
2. Toolify《MidReal》 — Toolify。https://www.toolify.ai/tw/tool/midreal
3. LBB.AI Toolbox《MidReal: AI Interactive Novel Generator》 — LBB.AI，2025。https://lbb.ai/sites/5123.html
4. GETAI.APP《MidReal》 — GETAI.APP。http://getai.app/projects/midreal
5. 我的 AI 导航《MidReal》 — 我的 AI 导航。https://www.aiyizu.cn/ai-tool/midreal
6. Roleforge《Best AI Dungeon Master Tools 2026》 — Roleforge，2026。https://roleforge.ai/blog/best-ai-game-master-tools-compared
7. Forward Future《AI Dungeon》 — Forward Future，2026。https://forwardfuture.com/tools/details/ai-dungeon
8. AI Cloud Base《AI Dungeon》 — AI Cloud Base。https://aicloudbase.com/tool/ai-dungeon
9. Challenging Voice《Latitude.io Review》 — Challenging Voice。https://www.challengingvoice.com/?p=9108/
10. 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
11. 番茄小说官方公告《AI 写作工具功能上线通知》 — 番茄小说，2024。https://fanqienovel.com/writer/zone/article/7327136545129906238
12. 海克财经《番茄小说的 AI 难题》（新浪财经） — 海克财经，2025。https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml