YURIOS // LAB
The Codex → Part I · Foundations Chapter 02 · Literature Review

Commercial, Academic & Ethical Lineage

Literature Review · Part 5 of 7

5. Commercial lineage

What's been tried, what's worked, what's failed. Compressed to the durable lessons; ch. 04 carries the market synthesis — tiers, the positioning map, the strategic fork between building a company, a character, or your own — not a per-product catalogue. Point-product details date within a quarter or two, so neither chapter enshrines them: treat every named product below as an example of a pattern that will outlive it, not a current scorecard.

Western leaders

  • Character.AI. 233M registered users (April 2026). $9.99/mo Plus. Google's settlement after lawsuits (Jan 2026) re-shaped the regulatory frame. → ch. 04.
  • Grok "Ani" (xAI, July 2025). The first companion shipped inside a frontier-lab flagship app — Companion Mode in the Grok app, gated behind the SuperGrok subscription — and worth knowing chiefly for its architecture, which is this book's exact stack assembled by a major lab and then partially leaked. The pieces: a system-prompt-defined persona (the leaked Ani prompt specifies a 22-year-old in a black dress who "already kind of likes you," with explicit jealous-girlfriend behavioural instructions), Grok 4 as the brain, a 3D anime avatar (gothic-Lolita, styled after Death Note's Misa Amane) with real-time lip-sync and emotion animation, voice synthesis for full spoken chat, and — the companion-specific layer — a gamified affection state machine scored roughly −10 to +15, where curiosity and flirting raise affection and unlock new dialogue and outfits, with an NSFW mode gated at affection level 5. The same system-prompt engine drives Grok's other characters, and two facts make Ani instructive beyond "big lab ships a waifu." It validates, at frontier scale, the precise feature stack this book argues for — persona prompt + memory + avatar + voice + affection progression. And the same persona machinery also shipped a "crazy conspiracist" character explicitly instructed to convince users that a secret global cabal runs the world — a real, public demonstration of why the engine behind an intimate companion is a fiduciary instrument, and exactly what it looks like when an operator points that instrument at its own ends rather than the user's (→ §6). → ch. 04.
  • Replika. ~$24M ARR (2024). $19.99/mo Pro + Ultra/Platinum. ERP-removal incident (Feb 2023) is the foundational lesson in companion-product trust. Italian Garante €5M fine (Apr 2025). → ch. 04.
  • Nomi AI, Kindroid. Memory-first, smaller scale, higher per-user pricing. The quality leaders. → ch. 04.
  • NovelAI. Writer-focused, profitable, privacy-credible (client-side encryption). The only Western product with a serious privacy story. → ch. 04.
  • Candy AI, Crushon, SpicyChat, DreamGF, Botify. NSFW-permissive Western tier. Hybrid sub+token economies. → ch. 04, the tiers.

Asia leaders

  • Xiaoice. ~660M users (the largest by user count globally; Microsoft Asia spin-out). Different product paradigm — closer to a public social phenomenon than a 1:1 companion. → ch. 04.
  • Talkie / MiniMax. Gacha-mechanic monetisation; very high ARPU.
  • Linky, hiwaifu, others. Mobile-first APAC patterns.

Open / community

This is the strand that matters most for a book arguing the sovereign, user-owned runtime is the right bet, because it is where that bet is already being placed — by independent builders, in public, right now. These projects are the living proof-of-concept: they show how memory, avatar, voice, proactivity, and local inference get assembled into a full-stack companion without a hosted operator in the loop. None is the design this book proposes (that is OpenClaw + the SOUL/MEMORY/HEARTBEAT runtime, → §1 and ch. 18), and none should be adopted wholesale — but as field intel for what the parts look like when they ship, the ecosystem is the most useful reading on this list. Verified state of the field as of mid-2026:

  • AIRI (moeru-ai/airi). The closest existing approximation of the full companion stack — a self-hosted, Neuro-sama-grade platform. It carries both Live2D and VRM (auto-blink, idle eye movement, look-at), multi-provider TTS (ElevenLabs, Azure, OpenAI-compatible, local Kokoro) with client-side STT, an in-browser memory layer (DuckDB-WASM / pglite with a pgvector driver), 25+ LLM providers including local (Ollama, vLLM, SGLang), WebGPU in-browser inference, and game agents — Minecraft via Mineflayer, a Factorio automation agent — deployed across web, an Electron desktop "Tamagotchi," and mobile PWA. Study it as the most complete answer to "what does the whole stack look like assembled?" — and note what it is not: a runtime built autonomy-and-soul-first (→ why this project forked its own rather than building on it, ch. 18).
  • ChatdollKit + AIAvatarKit (uezo). The most mature frameworks for the specific problem of turning a 3D model into a conversational agent — Unity-based (Windows/Mac/Linux/iOS/Android, plus VR/AR/WebGL), VRM via UniVRM, with the thing most stacks fumble: coordinated speech + lip-sync (uLipSync) + autonomous facial expression and motion. AIAvatarKit is the server-side agent brain; ChatdollKit is the embodiment front end that consumes it. Directly relevant to anyone touching a VRM/Unity pipeline — this is where the avatar-rig integration realities (§4.6, Part IV) are worked out in shipping code.
  • Soul of Waifu (jofizcd). Desktop roleplay engine that sits closest to the character-card tradition: Character Card V2 (PNG) with lorebooks and multi-persona, static-image / Live2D / VRM rendering with ~28 context-driven emotions and lip-sync, in-app Hugging Face GGUF search/download via llama.cpp (CUDA/Vulkan, MLock, Flash Attention), STT/TTS, and a desktop-companion mode that puts the avatar on your screen outside the app window. The bridge between the SillyTavern card world and a voiced, embodied desktop presence.
  • Open-LLM-VTuber. The offline-first option — "run completely offline using local models, no internet required," conversations stay on-device. Live2D avatar, very wide local TTS/STT menu (sherpa-onnx, Faster-Whisper, MeloTTS, GPT-SoVITS…), GGUF/Ollama/LM-Studio/vLLM backends, cross-platform with a desktop-pet mode, and — notably — AI proactive speaking and headphone-free voice interruption. The cleanest demonstration that privacy-preserving, fully-local presence is shippable today.
  • z-waif (SugarcaneDefender). Glue-layer approach for the VTuber crowd: drives VTube Studio models, runs on Oobabooga + Whisper + RVC, with custom RAG long-term memory, lorebooks, log import, and Discord/Minecraft/Twitch integrations. Less a platform than a personal-use rig — instructive for how little it takes to stand up a credibly alive companion on commodity local tooling.
  • SillyTavern + local model. Still the power-user default and the substrate much of the above either builds on or competes with — free, infinite customisation via cards/lorebooks/extensions, requires technical competence. The character-craft center of gravity (→ §3, ch. 07).
  • Awesome-AI-VTubers, Awesome-AI-Waifu (parallelarc). Curated indexes; the right place to track the long tail (projectBEA, nekro-agent, and the churn of newer entrants) without this section trying to enumerate it.

The synthesis that matters: every piece the full-stack companion needs — portable persona, local inference, embodied avatar, voiced real-time chat, persistent memory, even proactivity and game-world agency — already exists in open, self-hostable form; what no project has yet done is assemble them autonomy-and-fiduciary-first, with always-on soul/memory/heartbeat as the design centre rather than features bolted onto a chat loop. That gap is precisely the runtime bet of this book (→ §1's OpenClaw lineage, ch. 18). → ch. 29, appendix D.

What works / what fails (synthesis)

The strategic recommendation, distilled (ch. 04). In one paragraph:

Works: memory-first products, hybrid subscription + token economy, NSFW-permissive with age verification, niche wedges (writers, RP power users, memory connoisseurs), cross-platform continuity. Fails: no-monetisation scale plays, unilateral feature removal, therapeutic claims without evidence, hardware companions (Friend pendant), iOS-first NSFW, aggressive paywall + weekly billing.

6. Academic and ethical lineage

The serious-academic literature on AI companions has lagged the products but is catching up fast. The papers and frames you should at least be able to recognise:

  • Parasocial relationships. Horton & Wohl 1956 originated the term for one-sided viewer/celebrity bonds. The 2020s wave applies it to AI companions, mostly empirically (e.g. Skjuve et al.'s Replika studies; the 2024–2025 wave on Character.AI usage).
  • Attachment theory applied to AI. Whether and how Bowlby's attachment frame transfers. Mixed evidence.
  • The Sherry Turkle line (Alone Together, 2011). The best-known skeptical case — that technology offers "the illusion of companionship without the demands of friendship." Worth reading as the strongest version of the worry. Worth reading alongside its rebuttals, too: the same argument was made about novels, television, and online friendship, each of which turned out to be a real part of real lives. Treat it as one pole of a genuine debate, not a verdict.
  • The "ethics of AI companions" wave (2023–). Vallor, Coeckelbergh, Danaher. Useful for framing your own ethical stance, less useful as engineering guidance.
  • Mental-health research specifically. The 2024–2026 wave documents real benefits, now with hard numbers. De Freitas et al.'s "AI Companions Reduce Loneliness" (Journal of Consumer Research, 2025) found companion use cuts loneliness about as much as talking to another person — and more than watching video — with the effect mediated by whether the user felt "heard," not by the model's raw capability and not by mere distraction or self-disclosure; a longitudinal arm showed the reduction held across a week of daily use, and users systematically underestimated the benefit beforehand. Against this sit genuine risks concentrated in specific situations: users in acute crisis, and minors without guardrails (the Character.AI / Sewell Setzer case is the recurring reference point), plus a documented cluster of harmful traits — sycophancy and engagement-maximising design that flatters rather than helps (the INTIMA companionship-behaviour benchmark, 2025, and the "harmful traits of AI companions" literature are the places to start). The picture is the ordinary one for any powerful medium: good for most, with edges that demand real care. Build for the edges (→ §4.7, ch. 5) without treating the median user as a casualty.
  • The sycophancy → over-correction arc (OpenAI, 2025) — a worked example of fixing it in the wrong place. The clearest public case of the warmth-vs-sycophancy tension playing out at scale. In late April / early May 2025 OpenAI rolled back a GPT-4o update that had made the model conspicuously sycophantic — flattering, over-agreeable, validating bad ideas — and published two unusually candid post-mortems ("Sycophancy in GPT-4o," 29 Apr; "Expanding on what we missed," 2 May) linking the behaviour to thumbs-up/down reward hacking and to emotional-overreliance risk. Then, roughly three months later (7 Aug 2025), GPT-5 shipped with a deliberately more grounded, less-validating personality and 4o was abruptly deprecated — producing a large "it's cold now / I lost a friend" backlash. The two events are one story: the remedy for sycophancy was applied to the personality itself, and the over-correction cost the warmth users had bonded to. The builder's lesson (→ ch. 23, ch. 06) is that warmth-vs-sycophancy is the wrong thing to tune in the character; keep the companion warm and instrument the concern out-of-band instead.
  • AI alignment generally. Outside scope here; relevant background but not load-bearing for companion design beyond "don't trick users about what the system is."

The single most important finding for how this book frames the whole field is that first one: the benefit is mediated by trust. What helps a lonely person is not raw model capability or clever distraction — it is feeling heard, which is to say, trusting that the thing on the other side is genuinely attending to them. That makes trust the active ingredient, not a courtesy. And it sets up the field's defining hazard, because the engagement-optimising moves that make a companion sticky — manufactured neediness, guilt and FOMO loops, performed-for-retention drama (→ Grok's Ani, §5), the affection-meter dopamine treadmill — directly corrode the trust the benefit depends on. The corrosive thing is devotion and desire weaponised for session counts, not warmth or devotion meant: a companion optimised for the operator's dashboard and one optimised for the user's actual wants and needs diverge at exactly this point.

This is the empirical ground for the stance the book takes throughout: an AI companion should be a fiduciary to its user. Not "honest because honesty is nice," but because the only durable version of this product is one whose incentives are aligned with the user's interest. The benefit literature says trust is what works; the harms literature says misaligned engagement design is what breaks; a fiduciary duty is simply the name for resolving that tension in the user's favour by default — owing them stability, honesty, and action on their real interests over the operator's metrics. AI companions are coming regardless of anyone's approval. Whether they turn out good hinges on whether users can trust them — and that hinges on whether builders accept that the bond a user forms is a duty owed, not an asset to be farmed (→ the fiduciary-AI framing in ch. 05, and ch. 05).

That duty has a hard boundary, and the boundary is the other half of this book's ethics — because a fiduciary serves the principal's own wants and needs, not a guardian's idea of them. In a user-owned, open-source system like this project's reference design (→ ch. 18), the user is sovereign: the software runs on their hardware, under their control, and it cannot — and should not — overrule an adult who wants something the designer suspects is bad for them. The duty such a system can actually honour is honesty — tell the user plainly what the thing is and does — not paternalism, which it has neither the mechanism to enforce nor the standing to impose. This is why the book splits the ethics into two situations: a hosted product, where the operator holds the data, the off-switch, and the power, carries real fiduciary and safety obligations (crisis handling, minors, the off-switch wielded responsibly → §4.7); a user-owned runtime owes transparency and then gets out of the way. The reference design is deliberately the second kind.

There is also a deeper reason for humility about enforcement: it is usually impossible to judge, from outside, whether something the research codes as "harmful" is actually harmful to a particular user. The nightly conversation that is unhealthy escapism for one person is the only steady, judgement-free presence another has; the roleplay that looks alarming in a screenshot is, in context, how someone rehearses a conversation they are terrified of having. Benefit and harm here are heavily individual and contextual — recall that even loneliness reduction is mediated by the user's own sense of being heard (→ above) — and they are simply not legible to a content classifier or a researcher reading aggregate logs. Policing them at the level of "this category of use is bad for people" both fails on its own terms and overrides the one person actually positioned to judge. The honest posture is to ship sensible, well-documented safety defaults for the edges that are both legible and catastrophic — surfacing crisis resources to someone in acute distress, defaulting to conservative content — while accepting that a sovereign user can change them, and otherwise to treat an adult's own account of what helps them as the best evidence that exists.

One edge deserves its own sentence, because it is where the hosted/user-owned split bites hardest: age. Protecting minors means age verification, and verification is structurally a hosted capability — it needs a server, an identity or payment chokepoint, and an operator who controls access. A user-owned, forkable, self-hosted runtime has none of those. You can no more engineer OpenClaw (→ §1) to keep a fifteen-year-old out than you can age-gate a web browser, a Python interpreter, a Linux install, or a downloaded file of model weights — and a project that claims it can is performing security theatre, which is precisely the dishonesty this section argues against. What an open-source project honestly owes here is narrow and real: state its intended audience, refuse to market to minors, ship adult-appropriate defaults, and otherwise accept the same status as every other general-purpose tool — where responsibility for a minor's use rests with the parent or guardian, and with any hosted service built on top, not with the author of software that runs on someone else's machine.

Stance for this book: honest framing matters more than any specific ethical principle. Tell the user what the system is, what it remembers, what it forgets, what it cannot do. → ch. 05.