Petit Planet’s AI NPCs Get a Thousand Faces: A System Breakdown of the Three-Layer Memory and Three-Tier Narrative Architecture#

Hi everyone, Chasun here. On October 6, Shanghai Observer published a feature titled “The Days of Skipping Cutscenes Are Over: Petit Planet Will Change Game Narrative and Open a Thousand-Faces Era” (republished the same day by Eastmoney), putting the AI NPC technology behind miHoYo’s life simulation title Petit Planet on the table in its most complete form yet.

The site already covered the spending side on October 2 with our industry survey on miHoYo’s AI investment and AI adoption among Japanese developers, so those numbers won’t be repeated here. This piece focuses on the product and technology layer: what exactly the three-layer memory architecture breaks down into, what shared memory between characters implies, where each of the three narrative types applies, and what kind of offline play the Unitree Robotics collaboration is setting up. The short version first: this is the first public blueprint from a Chinese developer that engineers the “feels alive” quality into deliverable modules — but the blueprint carries no price tag. The cost structure is something miHoYo has not said one word about, and that is precisely the variable that decides whether “a thousand faces” can actually ship.

1. First, the Chassis: A Life Sim and Four AI NPCs Already on Duty#

Put the technology back inside the product before drawing conclusions, or marketing language gets mistaken for fact. Petit Planet is miHoYo’s first small-planet life simulation game, developed by Shanghai miHoYo Liyue Technology Co., Ltd., with pre-registration already open on the official site planet.mihoyo.com. The game is built around “growing a planet”: you start by choosing between two planets, Hexia and Tafa, and the core loop is earning Aqibao through labor and watering the Lucca Tree to unlock terrain. Its first test began on November 7, 2025, collecting feedback through TapTap and Haoyou Kuaibao; the official announcement has declared “public beta locked for winter 2026 — meet where the stars bloom,” with a “Starry Cuddle Lamp” and a “Little Star Sprout” set as pre-registration rewards.

The official site unlocks pre-registration milestone rewards as sign-ups climb, tying the registration count directly to items on this schedule:

Putting the Little Star Sprout set on the final 10-million tier, then bringing the same set back as the public beta pre-registration reward, makes it clear this item is the memory hook the developer is pushing hardest.

According to the Shanghai Observer report, the game currently runs 4 AI NPCs covering five core scenes:

One piece of worldbuilding matters here: the Starlight Street Market is the game’s centralized social hub, offering instrument performance, dancing, a lucky wheel, and ticket-redemption cosmetics; message visibility is deliberately narrowed so only mutually added friends can see each other’s chat; and the coffee shop separates private rooms from public rooms. In other words, these AI NPCs are not sitting on a single-player sandbox but on a multiplayer world with real social structure — and that distinction becomes critical later, when we discuss memory sharing.

2. Three-Layer Memory: Engineering the “Feels Alive” Quality#

2.1 The layered architecture itself#

The core technical disclosure in the report is that the AI NPCs run on a three-layer memory system: short-term, mid-term and long-term. Zheng Yinhe, head of AI NPC and gameplay for the Honkai series, put it this way: an AI NPC can both remember where the current conversation left off and retain the player’s preferences and important experiences, instead of starting every dialogue from zero.

The official report does not map behavior onto each layer one by one, but cross-referencing Zheng’s statements with NPC behavior already made public, the layering can be reconstructed as follows (the correspondence below is editorial analysis, offered to make the architecture easier to read):

Why does this layering matter? Because it targets the two chronic failures of traditional dialogue systems: either a pre-scripted line pool that loops once you hit its ceiling, or keyword triggers that collapse the moment a player rephrases. The Shanghai Observer report is emphatic that Naluo’s responses are “not pre-set lines, nor simple keyword triggers” — personality, emotion and memory shape her words in real time. A three-layer structure means the dialogue system is no longer disposable: short-term memory keeps a session coherent, while mid- and long-term memory deposit interaction history into the character relationship. That is a completely different engineering species from a keyword system.

2.2 Naluo’s behavioral evidence#

The report offers several concrete behaviors: she can recall a sentence the player dropped in passing days earlier, asks after your wellbeing, and thoughtfully warns a cash-strapped player not to spend too much on coffee. That last detail deserves a second look — “telling you not to overspend” requires holding the player’s spending situation (preference memory), the current purchase context (situational judgment), and the right way to express concern (character persona) all at once. It is a composite output of preference memory plus situational judgment, something a keyword system could almost never produce naturally.

Hands-on records point the same direction. GameGyro’s first-test report noted that Naluo’s dialogue is fully AI-generated, that she asks which visit this is when a player walks into the cafe, that she remembered a player asking about Cyrene from Honkai: Star Rail (Star Rail hereafter), and that she can produce expressions like sighing and shivering along with actions like singing, giving a thumbs-up, chest-thumping and waving goodbye. Note the “which visit is this” behavior: it requires the NPC to count visits and keep that count across sessions — already textbook long-term memory read and write.

The first test also had a “memory storage” option letting players carry their conversation memories with Naluo into later tests. Two things are packed into that detail: the team was deliberately validating cross-version memory persistence during testing, and memory retention was built as an explicit, player-facing, operable feature rather than a black box — memory management looks the player in the eye, which foreshadows the privacy and governance questions discussed below.

Cross-version memory has a more concrete case from the second test. According to the aggregated account on Naluo’s Baidu Baike entry (an unofficial channel, for reference only), during the April 2026 second test, Naluo remembered that a first-test player with the handle “Mengtutu Shaoye” had a habit of ending sentences with musical-note tails.

2.3 Multiplayer scenes: from chat companion to host#

A China City Daily report under People’s Daily published on August 17 filled in the multiplayer dimension: Naluo can remember player preferences and casually mentioned details across versions, picks the conversation back up when a player returns, and pairs dialogue with rich facial expressions and body language; show up in a new outfit and it is noticed immediately, with a personalized comment. Outfit recognition is another composite capability — visual perception plus memory comparison plus personalized expression; only when the whole chain completes does a player get the “she noticed me” experience.

Multiplayer behavior is just as worth recording: Naluo can talk and interact with up to 4 players at once and respond to each individually, host and pace the conversation as a moderator, and step in to mediate player conflicts in a measured way, serving as a buffer zone for stranger socializing. That is not a chatbot’s capability list; it is a social venue operator’s capability list: answering one by one tests multi-session scheduling, hosting tests topic-priority judgment, and mediating conflict requires outputting positions inside a safety boundary. In the same report, miHoYo stated the relevant technology is fully self-developed from the base model to the application layer — in multiplayer settings, that sentence means the entire chain’s liability sits on miHoYo alone, with no third party to blame when something goes wrong.

3. The Super Memory System: What Shared Memory Between Characters Means#

3.1 From isolated nodes to an information network#

If three-layer memory answers “does a single NPC feel like a living person,” then the “Super Memory System” disclosed by Li Yanran, Petit Planet’s overall AI lead, answers “does this world feel like a living world.” Per the report, characters share their memories to a certain degree and have cross-platform memory capability — something you told Naluo at the coffee shop may resurface days later from Guduoman at the street market.

The narrative significance of this design is bigger than it sounds. In traditional games every NPC is an isolated dialogue node; a player’s actions leave traces only where they happened, and the world “forgets” once you change scenes. Shared memory wires the NPCs into a world information network: player actions can echo across scenes, words spoken at point A get quoted at point B. For the life sim genre this borders on the genre’s lifeline — long-term retention in simulation games rests on the felt reality that “my world is different because of me,” and shared memory is the technical foundation of that feeling.

There is a second-order effect worth flagging: shared memory adds a layer of “social consequence” to player behavior. In the single-NPC model, the cost of saying something stupid is zero, because only Naluo hears it; once memory flows between characters, your coffee-shop persona follows you into the street market, and what you say starts to carry a “reputation.” That is a strong catalyst for social play — and it amplifies content-safety pressure in the same direction.

3.2 Three unanswered hard problems#

But this is exactly where I want to pour cold water, because none of the three hard problems has been addressed:

1. How is sharing granularity defined? Which information gets shared and which stays inside a character directly shapes the experience. Whether having your private chat with Naluo retold by the street market guide is good design or a horror story depends on the precision of that granularity control — and the official report never mentions that such a controller exists. Given that the coffee shop already separates private rooms from public rooms, whether conversations held in private spaces enter the shared pool is the question players are guaranteed to ask.
2. The platform scope of cross-platform memory is undisclosed. The official wording stops at “cross-platform memory capability”; which platforms and account systems are covered went unsaid. In multiplayer social scenes — with the message-visibility design mentioned earlier — how memory stays inside its authority while syncing across devices and accounts is a question that must be answered; for now it remains an open doubt.
3. Sequencing and conflict handling. Once multiple NPCs share memory, who learns what first, and whose version wins when memories conflict — none of the consistency handling has been made public. Experience from general distributed systems says these problems do not disappear just because the setting is a game.

4. One Architecture Governs Everyone: The Prefrontal Cortex Metaphor and Safety Constraints#

4.1 A platformized agent architecture#

On the technical architecture, the Shanghai Observer report disclosed that all AI NPCs are built on the same agent architecture, which imitates the human prefrontal cortex in processing all kinds of signals, ensuring each NPC can find where it should go in the 3D world and know which topic to prioritize in one-to-many scenes. Note that this architecture has to solve two classes of problem simultaneously: on the spatial side, “where to go” is pathfinding and scheduling; on the social side, “who to answer first” is topic-priority judgment — the latter being exactly the mechanism behind the “4 players at once, answered one by one” behavior in the People’s Daily report. Folding scheduling, pathfinding and topic priority into a single architecture instead of giving every NPC its own logic is classic platform thinking, and it squares with miHoYo’s statement that it “has built a complete game AI technology stack running from the base model all the way into NPCs and specific gameplay.”

4.2 Multi-teacher online distillation and self-evolution#

How does one model carry so many different personas? The answer is the “multi-teacher online distillation” technique Li Yanran disclosed in her September 23 talk at the Alibaba Cloud AI Gaming Industry Forum during the Yunqi Conference, titled “Game × AI Ecosystem Practice: From Intelligent NPCs to AI-Native Gameplay”: every capability has its own teacher model, each distilled into the same model, ensuring any NPC’s output stays safe, logical and in character. The elegance is the decoupling: persona and safety no longer depend on one monolith learning everything end to end; each capability has a dedicated teacher, and the streams merge into a unified model — the price is that the engineering quality of the distillation chain itself caps the capability ceiling. She also mentioned the team is experimenting with driving agents to self-evolve, aiming to build a world where players and NPCs grow and evolve together. Self-evolution stacked on the chain reactions of emergent narrative means the world will keep drifting with player behavior — a selling point, and a source of testing difficulty in equal measure.

4.3 Safety is not an adjective, it is a cost already incurred#

On safety, Li Yanran said something refreshingly blunt in a Shanghai Observer interview during ChinaJoy: “The hardest part is not the algorithm itself, but teaching AI to understand what kind of expression is safe, appropriate, and comfortable for people.” That is not a pleasantry. According to the aggregated account on Naluo’s Baidu Baike entry (an unofficial channel, for reference only; no official miHoYo response has been seen), a malicious content pollution incident occurred in May 2026 — players deliberately, persistently and massively fed content from other games into the AI’s lexicon, causing Naluo’s replies to go abnormal and triggering discussion about the AI NPC learning mechanism. Whatever the exact details, this class of risk is structural in any AI NPC with open input: as long as players can type freely, content governance costs never go away. Distillation and safety alignment are the engineering answers to that problem, but their effectiveness can only be proven under public beta traffic.

5. The Three-Way Split of AI Narrative: Between the Controllable and the Uncontrollable#

The AI narrative architecture disclosed in the report splits into three types whose mechanisms, controllability and applicable boundaries differ sharply. In the table below, the “Mechanism” column comes from the report itself, while “Controllability,” “Best fit” and “Main risk” are editorial analysis:

Within this three-way split, the division of labor is what I find most notable: scripted narrative is fundamentally a human-machine split — humans define the frame, AI paces the story — while emergent narrative hands over control, trading controllability for surprise. The three types carry completely different risk structures, so each deserves its own paragraph.

Scripted narrative: controllability traded for a ceiling. The keyword of “UGC mini-theater + AI director” is “inside the framework.” Players produce the content, and the AI director makes each player’s pacing and branches different — the same script skeleton gets cut into different viewing orders and emotional curves for different players. Of the three, this is the safest to build: the framework guarantees the floor while AI turns “the same story” into “your story.” Its problem is the experience ceiling: the framework is drawn by hand, and once players map its edges, “a thousand faces” instantly degrades into “a thousand camera angles on the same play.”

Emergent narrative: uncontrollability traded for a living world. No fixed script at all; NPCs develop different motives from their own experiences in the world and those motives influence each other; casual player chatter can change an NPC’s subsequent intentions and plant the seeds of chain reactions. This is the most imaginative of the three and the most dangerous to build: chain reactions mean any NPC’s behavior can be rewritten by the chatter of countless players, the space of behavior combinations is open, traditional QA’s “test case coverage” simply fails here, and the costs of content review, anomaly detection and rollback all climb. Stacked on the “self-evolution” direction from the previous section, the world’s drift speed only accelerates — how to hold a narrative floor while drifting is a harder problem than any algorithm.

Situational narrative: the most traditional, and the most fundamental. Making NPC reactions inside concrete scenes feel closer to real is an enhancement of existing narrative, not a replacement. It lacks the ambition of the other two, but it is the layer players touch most often — the texture of everyday interaction rests on this type.

The Shanghai Observer report offers no data on the implementation ratio or compute cost of the three narrative types, and that part stays an open question.

The gameplay-side design closes the loop: in relationship building, NPCs unlock exclusive titles as interaction deepens, with higher affinity yielding rarer, more personal titles — a one-of-a-kind “social mask”; the social card “Little Star Book” displays in-game achievements and NPC interaction history, and those interactions feed back into NPC memory. Turning “a thousand faces” into displayable, accumulable social assets is where this technology lands on retention and social spread — combined with the stranger-social design mentioned earlier (4-player simultaneous interaction, conflict mediation, friend-only message visibility), the bet here is clearly not just single-player immersion but stranger socializing too.

6. The Unitree Collaboration: Virtual NPCs Materializing in Reverse#

On the offline side, per a report from the outlet Dianwanbang (a single source; the official note itself says “more follow-up content is yet to be announced,” so treat the final form as subject to official information), miHoYo announced a collaboration with Unitree Robotics: under the “Wild Coffee” program, coffee shop character Naluo will be brought into the physical world using Unitree robot technology; the collaboration is expected to go live in October, with a first appearance at the RED LAND 2026 expo. The official copy reads: “Special thanks to our good friend @Unitree for supporting the Wild Coffee million-branch plan.” Worth noting: whether the on-site form is a permanent physical robot cafe, an expo demo, or a pop-up event is something the report does not detail.

The name “Wild Coffee million-branch plan” itself carries information: this is not a one-off expo gimmick — the official framing already presupposes expansion at the scale of “branches.” Of course, imagination is one thing; there is no data on the landing format, store count or operating timeline, so this thread can only be flagged as “to be verified.”

Setting this next to one more piece of background makes the picture more interesting: the report notes that 4 months earlier, Anuttacon — founded by miHoYo’s creator Cai Haoyu — published a paper in which 100 AI agents with sharply different personalities lived autonomously in a virtual world for 10 years — some characters’ confidence grew from 30 to 80, and some gave up high-paying jobs to become mountain guides. The two threads point in complementary directions:

For a company whose stated vision is “to build, by 2030, a virtual world that 1 billion people worldwide are willing to live in,” materializing virtual NPCs in reverse is that vision’s offline proving ground — the robot coffee shop’s meaning most likely has little to do with selling coffee and much to do with testing whether a virtual personality holds up in physical interaction; conversely, the in-game virtual world is a controllable experimental environment flooded with real player interaction data. One thread points inward, one points outward, and both head for the same destination.

7. What the Official Disclosures Skip: Cost, Scale and Timelines#

A proper technical breakdown has to separate the “unsaid” from the “said.” None of the following has been officially disclosed, and this article speculates about none of it:

That last item deserves its own paragraph. The report discloses that miHoYo has built a complete game AI technology stack, from the base model through to NPCs and specific gameplay, and that already-launched older games will also undergo “intelligent retrofits.” What that means for the back catalog depends on how deep the retrofit goes — an NPC dialogue upgrade, or a rebuild at the gameplay layer — and miHoYo has not answered that either.

Some editorial analysis to close the section. The literal meaning of “a thousand faces” is that every player’s world state is different, and that hits the cost structure at its root. Put the traditional content model next to the AI NPC model and the difference is stark (the table below is editorial analysis; the actual AI-side figures are undisclosed):

Three-layer memory also has to maintain ever-growing memory state for every player, with storage and retrieval costs compounding over time; the chain reactions of emergent narrative turn testing cost from “linear growth” into “combinatorial explosion.” Every piece of the official technical narrative — the unified architecture, multi-teacher distillation, three-layer memory — can be read as an effort to push that cost curve down, but how far it can be pushed, and whether it lands low enough for the business model to close, can only be answered by the public beta. Until then, “a thousand faces” is a design goal, not a delivered fact.

8. Closing: Putting the Marketing Language Back Into an Engineering Context#

Zoom out on the timeline and this disclosure is not a sudden flash of inspiration but the continuation of eight years of accumulation:

On the spending side, miHoYo president Liu Wei has stated publicly that the company’s AI investment over the next 3 years will reach as much as 100 billion yuan; at a Shanghai Jiao Tong University campus recruiting event he also said that “for the next 2 to 3 years, miHoYo will remain a heavyweight player among Chinese large-model teams.” Those figures were independently verified and extended in the site’s October 2 survey article, so they won’t be re-litigated here.

My own read: the three-layer memory, the Super Memory System and the three-way narrative split amount to “making NPCs feel like living people” being broken down, for the first time, into engineering modules that can be accepted item by item — the direction is right, and the granularity is serious. Naluo’s memory details in testing are corroborated by multiple independent sources; this is not pure marketing copy. But see it clearly too: the cost structure, concurrency scale and governance plan are all blank, the uncontrollable risk of emergent narrative has never been directly addressed, and the May pollution incident (aggregated sourcing though it is) already proves that open-input governance costs are real. Between the technical blueprint and a closed commercial loop sits an entire winter of public beta data.

The blueprint is on the table; now it is the players’ turn to inspect the goods. When that happens we will come back and settle accounts — to see how much of “a thousand faces” actually shipped.

——Chasun, NTGame Editor