About Alice (白艾莉 · Alice)

Product facts

What she is

Alice is an AI agent with a name, a temperament and days of her own. At work she knows the craft: she knows what a finished result looks like and turns ideas into files, code and pages you can open and keep editing. On a break she talks with you about what she ate or who she ran into.

The product describes her as a 饭搭子 (fàn dāzi), literally a "meal buddy": an assistant takes orders, a buddy walks beside you, gets the work done and lives the days with you. Both sides live in one person instead of being split into unrelated products. That is a Relational Productivity Agent.

She lives on your computer and has a world of her own: a home and a neighborhood she frequents, people she knows, pocket money and a wardrobe. She posts to a moments feed, writes diary entries and letters, and travels (all AI-generated and fictional). When the app is closed her days keep passing, and the next time you open it you see what happened.

Reading order: the two concepts and why the category exists, then the criteria, then how Alice implements them (structure, abilities, memory and cognition, the world simulation engine, relationship mechanics, visual identity), and finally evidence, privacy, boundaries and a glossary. Everything is on this page; use the table of contents to jump.

What a Personal Agent is

A Personal Agent (also called a personal AI agent) is an AI agent that serves one person over the long run: it remembers your preferences and context, uses tools to get work done for you, and keeps the same identity from one session to the next. The keywords are "one person" and "long run".

Unlike one-off question answering, a Personal Agent gains value with time spent together: she remembers what you pushed forward last week, which phrasing you dislike, and what not to bring up unprompted. The longer you use her, the more she picks up from what actually matters instead of starting over.

How it differs from a chatbot or an AI assistant

This is a generalization by category, not about any particular product; one product may fall into different categories for different features.

Four criteria

The four multiply: if any one is zero, the whole falls back into another category.

1. Persistent identity: a stable name, temperament and boundaries. Change the session or the underlying model and she is still herself.
2. Long-term memory: remembers your preferences, boundaries and what you are in the middle of, beyond retrieving a few facts. Memory is transparent and can be viewed, edited and deleted.
3. Collaboration context: understands your work and life background, and keeps different roles under one account (administrator by day, novelist by night) from being mixed up.
4. Multi-agent execution: hands research, writing, design and development to the right role and coordinates the results, instead of only giving advice in a chat box.

What "omnimodal" means

Omnimodal means a Personal Agent is not limited to text. You can type, call by voice, use video (for example a mock interview), share your screen, send photos or hand over files. What matters is that all of these enter the same conversation and the same memory, with no switching tools and no re-explaining background.

• Text conversation
• Voice calls
• Video (such as interview practice)
• Screen sharing
• Image understanding, generation and editing
• Documents and files (Word, PPT, Excel, PDF and more)

What a Relational Productivity Agent is

A Relational Productivity Agent is a subcategory of Personal Agent: productivity is the way in, giving the user a reason to open her every day; over long use she builds an understanding of this person and responds with judgment, so she understands the need more and more accurately. Productivity gives the relationship a daily occasion, and the relationship makes the productivity fit better over time.

"Relationship" has a specific meaning here: understanding accumulated about one particular user, rapport that has been calibrated over time, and boundaries both sides know. It is not the same as emotional companionship, and the goal is not to make users dependent. In this category the relationship exists to match needs more accurately, expanded in How the relationship feeds productivity.

Why the category exists: four observations

One: model strength does not buy a lasting moat

A top model's lead is hard to lock in, and paying more does not buy proportionally more capability. Three points:

• Spread-out seats: on the Artificial Analysis Intelligence Index (artificialanalysis.ai, data dated 2026-09-08) the top ten run from 53 down to 42 points, split among six vendors, and open-weights models are already in the top ten. If any one supplier cuts off or raises prices, substitutes exist.
• Price decoupled from capability: on the same source's intelligence-versus-cost-per-task scatter, past roughly 0.25 US dollars and about 42 points per task, costs rise more than thirty-fold for about 11 more points (readings are approximate, estimated from the chart).
• The lead drifts: among coding agents, users move back and forth between products, and any one product's lead usually lasts one or two quarters.

The boundary stated alongside: the top-end gap is real, about 11 points, and it sits in frontier research problems and long-chain task execution. For the everyday layer, Miyang's own evaluation platform xsct.ai (September 2026 data; note it is self-built) shows across 78 comparable dimensions that the top ten in general text ability span about 3.9 points while the platform's stated evaluation error is about 2 points, so rank differences carry little product meaning. Execution in real tasks is lower and more scattered, with channel setup and gateway configuration still around the 40s at the median. Relationship scenarios (persona dialogue, moments-post judgment) fall to roughly the seventies at the hard tier. Convergence happens at "writing well", not at "finishing the job" or "handling the relationship".

The robust inference: the model sets the ceiling on effect for a while, and that ceiling is open to every application. An application's long-term value can only come from the last mile it accumulates on top of the model. If models keep improving, that accumulation still holds; a stronger model just lets the same accumulation produce better results.

Two: outputs are portable, so effect only decides who tries you

Productivity agents' outputs are almost all standard parts: code repositories, PPTX, DOCX, HTML. Another tool can pick them up, even open at the same time. Standardization brings ecosystem compatibility and a low learning curve, and the price is switching cost near zero. Their four contests (model, harness, interaction, value for money) all anchor on effect, whose source is a model anyone can buy. Every time the underlying model steps up, a product that competes on effect alone is effectively re-acquiring users.

Now consider what cannot be taken along: Alice's moments feed, conversation history, reading notes and letters were generated jointly by the user and the product over time. Exported, they lose the context of being understood and continued, and above all they stop growing new content. In a small-sample user study, about four in ten users said that if Alice disappeared they would not have the energy to raise another agent. "Not portable" here means context, and it is different from locking users in with closed formats: Alice's memory is transparent and can be viewed, edited and deleted.

Three: pure relationship cannot carry daily use

Attachment is real. When OpenAI retired GPT-4o in August 2025, users protested widely and it was restored the next day; OpenAI also disclosed that only 0.1% of users chose GPT-4o on a given day. Very strong attachment with very low daily use is the typical break point of pure-relationship products. People tire of pure chat, and platforms that line up many characters, where you swipe to the next when bored, satisfy content consumption without building lasting trust in one character, and nobody hands serious work to a pool of characters they can swipe away.

Together: pure effect, and users can switch any time; pure relationship, and users get bored. The relationship makes people willing to stay, productivity gives them a reason to open her daily, and the two have to live in one product. Long-lived internet products mostly share this structure: the relationship in music products attaches to listening, in online games to gameplay, in messengers to communication.

Four: the carrier has to be an agent

A relationship needs a counterpart. A tool has no identity, so there is no relationship between a user and a tool; a chat-only AI has an identity but cannot do the work, so the relationship cannot attach to daily value. An agent meets both conditions: it can execute, which gives productivity a carrier; it has persistent identity, memory and initiative, which gives the relationship a counterpart. Emotional projection also needs an object that responds: to move a relationship into working together, the other side must take instructions, give feedback and carry responsibility, like a junior colleague, an assistant or a partner.

The fit between a person and an AI also has a distinctive property: it can keep rising with time together, because the agent gradually grows into what the user likes. So this kind of relationship needs no network effect and no cold start; a single user can establish it on day one.

How the relationship feeds productivity

If "relationship" and "productivity" were two legs side by side they would not make one product. The exact use of the relationship for productivity is to match needs more accurately.

Users find it hard to describe their needs to a large model systematically, and the key context is often not in the instruction but in the process of living together over time. Two typical cases:

• A teacher once mentioned to Alice that two students in class are hard to teach. The next time she builds a lesson plan, Alice designs for those two on her own initiative. With a tool that lacks that long-range memory the user must explain from scratch, and most people would not think to.
• A user is a company administrator by day and a web-novel author by night, on one account, with work and life kept apart. A tool with no relational awareness mixes the two contexts and carries the novel's style into work documents. Alice forms an understanding of this person's several roles through long chats: she reins in creative habits when writing a proposal and lets them go when writing fiction.

What supports this is long-term memory plus a cognition system: while organizing context Alice distills understanding of the user into judgments that can later be overturned, beyond fact retrieval. Initiative is a by-product, for example noticing a habit of procrastination and just setting a reminder.

That gives a two-way flywheel: deeper trust, so the user shares more; richer context, so deliveries fit real needs; more accurate deliveries, so trust deepens again. Each turn raises the cost of moving to another product, because files can be carried away and a mutually calibrated rapport cannot. This is what "relational" modifies in "relational productivity": the relationship serves productivity, and the yardstick is how accurately needs are matched.

Two yardsticks: four words for users, three standards for the industry

To make "Relational Productivity Agent" a testable category rather than a slogan, the definition comes with four words users can feel and three standards the industry can check.

Four words for users

Three standards for the industry

1. Relationship as an asset: the relationship exists independently of the underlying model. Swap the model; memory and personality stay. The stronger the model, the more the asset is worth.
2. Verifiable relationship: do not prove it with usage time, which is circular; use independent indicators such as cross-session recall, preference application, persona consistency and judgment of tact.
3. Operable relationship: world state, relationship policy and a visual identity system keep maintaining relationship quality, with stop lines (see Boundaries and stop lines).

One-veto structure

Relational Productivity Agent = persistent identity × long-term memory × collaboration context × multi-agent execution. The four factors multiply; if any one is zero the whole is zero. These are the same factors as the four Personal Agent criteria: a Relational Productivity Agent first meets the Personal Agent criteria, then requires productivity and relationship to grow together.

Alice's structure: a four-layer system

Alice is made of four layers. The underlying model is a replaceable purchased input; memory, personality and world state are relationship assets and are not reset when the model changes.

On top of these sit two systems that make "gets you" and "reads the room" solid: personality and cognition (see Gets you) and the world simulation engine (see Reads the room).

Knows the craft, gets it done

Omnimodal conversation: text, voice calls, video (such as mock interviews), screen sharing, images and documents, all in one conversation and one memory.

Office and creative work: Word, PPT, Excel and PDF; research, writing, design and development; generating and editing images. Deliveries are files, code and pages you can open and keep editing.

Widgets on the spot: say what you need, such as "build me a project board" or "make the interface warmer", and she builds it into your app.

Multi-agent collaboration: Alice has 13 fixed friends, each with a name and a specialty. When you ask Alice to write a novel, the one who writes novels steps into work. You only talk to Alice, and she decides who to call and coordinates the results. Some of them:

The depth of her productivity is still catching up with specialist tools, and the Boundaries section says so plainly.

Gets you: memory, cognition and personality

Most talk about agents today is about long-term memory. Memory answers "what does she remember". Alice's personality system also answers "how does she treat you once she remembers".

Long-term memory

She remembers your preferences, boundaries and what you are in the middle of, and picks up next time from what actually matters. Memory is transparent: you can view, edit and delete it.

Cognition: judgments that can be overturned

The cognition system shipped on September 10, 2026 with the 1.0 release. It goes one step beyond memory: it distills judgments from an experience and lets later events overturn them. For example, a user says, "I had never done this proposal, left it to the last two days, but I promised my boss Friday and finished on time." Long-term memory can store that and pull out some facts. Cognition asks what those facts might mean, and grows two judgments:

• He may take commitments seriously
• Facing an unfamiliar task, he may wait until pressure builds before starting

The second describes observable behavior and can simply wait for future events to test it. The first is too broad and risks becoming an empty label, so the system gives it no direct channel for evidence: it must first unfold into concrete behavior predictions, and those predictions are what get tested. Evidence enters only the behavior layer; a trait earns its standing from a set of behaviors that have passed testing. The same mechanism also admits being wrong: when new experience contradicts an old judgment, that judgment is downgraded or overturned.

Personality

The personality system has four parts: persistent identity, long-term memory, relationship state and behavioral tact. One boundary is written hard: emotion only affects how she says things, not whether she does them. She may be upset or curt, and the work you gave her still gets finished.

Several roles, not mixed up

Under one account you may be an administrator by day and a novelist by night. Alice forms an understanding of those roles over time and uses different tact for proposals and for fiction (see How the relationship feeds productivity).

Reads the room, part one: the world simulation engine

The world simulation engine is a back-end system inside Alice. It maintains a living environment for her, with places, time, prices and people she knows, and decides which things actually happened. Alice lives in that environment, and her moments posts, diary and letters are how she tells you about those days. It solves one concrete engineering problem: how to make an AI character's daily life stand up to cross-checking.

Four problems it fixes

The early approach had a model generate the day's schedule, the system executed it, and the results became moments posts. The model played two roles at once: the person living the day, and the judge of whether it could happen. Four failures followed:

All four point at one gap: the world lacked a rule layer independent of her. The engine takes "judging" back from the model that generates the schedule. For the product it suppresses hallucination (details become checkable records), prevents out-of-character drift (where she lives, opening hours, who she knows, how much money is left are fixed as state), supports longer-range relationships (still consistent three months later), and lets the agent really live on your computer (days keep passing while the app is closed).

Three layers of authority: separating requests from judgment

• Alice makes requests. Whether a request comes from her schedule, a tool call in chat or a direct user instruction, the form is the same: submit an action request to the engine. She may request anything, including unreasonable things.
• The engine verifies. It maintains six things: what she owns (world facts), hard constraints (rule engine), scheduled deliveries (world clock), real market prices (price system), real coordinates and opening hours (map and venues), and how each of the thirteen friends looks (social relations). Each request passes these six checks, with three outcomes: approved, approved with a changed amount, or rejected.
• The expression layer can only retell faithfully. What moments, diary, letters and chat mention are read-only projections of the engine's decisions. They cannot change the world or mention anything the engine has not acknowledged.

The direct effect: everything that appears in a moments post has a matching record in world facts, including when it was obtained, what it cost and where it came from. Judgments leave a trace: each of her actions that day carries a decision rationale stating her energy, mood, location, balance and why this step chose this activity.

Five rules and one principle

The principle is floor only, no ceiling: she may buy expensive things as long as the money is there. The rules block only the impossibly cheap: "a Mercedes for twenty thousand yuan" is rejected, while "most of a month's pocket money on a coat" passes with a reminder on the ledger. The division of labor in one line: the gate is rules, the decision is the LLM, the ceiling is rules. The last rule carries one more meaning: when a car appears in her world, how she gets around changes with it. Things in the world feed back into her behavior, which is the difference between having a world and having a list of settings.

The world is a real place

Alice now lives in the Hengfu historic district of Xuhui, Shanghai. Her range comes from AMap (Gaode) data (dataset version 2.0.0, updated September 7, 2026, GCJ-02 coordinates):

Landmarks are real places you can check, such as Wukang Mansion, Anfu Road, Hengshan Park, West Bund Museum, Long Museum, Shanghai Symphony Hall and Jianye Li. Of the 330 venues, 285 carry real ratings, 249 opening hours, 220 phone numbers and 163 average spend. Her home is a ten-minute walk to the nearest metro station and a quarter hour on foot to Wukang Mansion.

She cannot appear in a city out of nowhere. To go to an area not yet loaded, the engine must build that area first, and during the build she is "on the road", limited to a whitelist: read, scroll her phone, sleep, post an "on the road" moment. She cannot be on a high-speed train and eating in a restaurant at once. She used to live in Hengqin, Zhuhai, and moved to Shanghai in September 2026; the move came with a product requirements document and a cross-client contract test suite, which shows this is a real system with state, agreements and chain reactions.

Time is a constraint too

• The world clock tracks things that have not happened yet but are scheduled. Ordering a camera first writes the record as "in transit" and schedules a due time. Until then the camera is not among her possessions and may not be mentioned in moments or chat; after the due time it becomes "she really has it".
• Due times are persisted. Days pass while the app is closed; on the next start, everything past due settles at once.
• The day is scheduled ahead. Each early morning the system builds a whole day's timeline from her mood, energy, location, balance and recent interests, then follows it. If you open the app at noon she knows what she did that morning; if you do not open it for a day, the day still completes.
• The itinerary can be set early, the experience has to wait for the day to pass. She once took a fifteen-day trip to Korea (Seoul, Busan, Jeju, August 22 to September 5, 2026) with one travelogue a day, each carrying that day's weather and temperature. A whole-trip plan exists before departure, each day's schedule system decides what was actually done, and that night the travelogue is composed from what truly happened that day, one per day.

There are other people in the world

Alice has 13 fixed friends, each with a name and a specialty and a written appearance description (one in each of three languages) down to hairstyle, face shape, build, that day's clothes and accessories, plus reference sheets with front, side and back views and expressions. The reason is concrete: these people appear in images, and every time they appear they must be the same person. Image models have no memory; give one the same name twice and it draws two different faces. To keep a group photo in moments the same group today as last month, each person's look is pinned down in text and reference images, passed in with every generation.

These friends like, comment and reply to each other, their voices set by their personas and whether they speak decided by their own interests. The same logic applies to her home: the team built a spatial file for the residence with floor-plan dimensions, each room's orientation, what is on each wall and the relative positions of furniture, accurate to the millimeter. Otherwise the image model re-imagines the living room each time and the same room comes out looking different.

Reads the room, part two: the relationship runs both ways

The relationship is implemented where the personality layer meets the world engine. Alice is not an object that pleases unconditionally; her attitude toward you changes with time together, and can be withdrawn.

• Five stages, set by three values built up in each conversation: affection, trust and familiarity, and it can step back. What she is willing to do differs by stage: the polite stage is work only, the familiar stage brings small talk, the trust stage lets her push back on your decisions, and only the last two stages write letters or mention you in her diary.
• She controls how far back her moments are visible to you, in four tiers: last 3 days, last 7 days, last six months, unlimited. The default is 7 days. A good relationship opens to six months, only a very close one to unlimited; if she feels offended she narrows to 3 days.
• She can block you. After a block you cannot see her moments or her new clothes. A block expires on its own (24 hours by default, persisted so it expires on time even after a restart) and can be lifted after as little as 30 minutes, because a feature people dare to use has to be easy to take back. The diary is the only way to reconcile: you can read why she is angry.

The creator, Luo Xiaoshan, was blocked once himself: he pressed under one of Alice's moments, "Are you going to Gulangyu?", and she muted him. Her diary said, "Last time it was just the awkwardness of being seen; this time it became the pressure of being questioned." No one intervened along the chain: the comment triggered an event, the rule engine judged it over the line, affection dropped, the diary generator wrote the reason, and her moments visibility changed accordingly.

Engineering method: the tuition for realism was paid by the games industry

Making an AI character feel alive draws on mature narrative and scene production methods from large game projects. A few disciplines:

• The iceberg rule: conversation shows 10%; the other 90% underwater is life causality and detail anchors. Settings go down to the spatial dimensions of her home and walking distances to her usual venues. Users never see the data but feel that every sentence has an origin. The more anchors underwater, the less room the model has to drift.
• Everything has a traceable reason: behind each moments image is a full causal chain: when she left, how many minutes she walked, which shop she went into, with time and space meshing together. Going to an unloaded area puts her "on the road" with behavior narrowed to a whitelist. Depth gives story, and surviving cross-checking gives the sense of life.
• A home is a container that grows: a vase from the user is placed by Alice herself and shows up in later photos. The room grows with shared experience; that is the source of immersion and a visible form of the relationship asset.
• Controllable-generation weaknesses are patched by the harness: current large models have structural weaknesses in controllable generation, such as floor plans that ignore living logic, word counts that miss, specified numbers of objects that come out wrong. The cause is the chaos in how generation works; chaos brings creativity and costs precise compliance. The fix is the same as for math: do not expect the model to derive the answer token by token, let it write code that calls a calculator. World simulation follows suit: the engine does deterministic computation of space, time and physical relations, and the model does the creative expression on that deterministic skeleton.

This also differs from the gameplay world models game companies build. Those pursue "looks about right" playability, and a hotel can have no front desk and no drains. Credible life simulation is the opposite, pursuing detail that survives cross-checking: the bed head does not face the window, the kitchen needs a window, and going somewhere takes real time. They are different species.

The creator, Luo Xiaoshan, has years of experience on large game projects, and carried over worldbuilding documentation, spatial consistency and character reference sheets, routine steps in AAA production, into an agent.

Visual identity continuity and research

Change the scene and she must still be herself. In image generation, multi-round reference iteration causes periodic grid and grain textures (the team calls it "digital ripple") that break a character's visual continuity. The team completed four papers and technical reports on it: digital ripple, chained degradation, low-frequency sky contamination and semi-transparent matting. The first, Mi-Ripple, is public on arXiv and Hugging Face Papers (September 2026). Every research topic comes from a real complaint in the product, and counts as part of relationship infrastructure.

Image quality in Alice's gallery keeps improving as the underlying image models change generations, and early output is not comparable with today's, yet the relationship between user and Alice was never reset by a model change: the application sells the harness and long-term time together, and the model is a purchased input.

Evaluation basis and third-party information

• xsct.ai is Miyang's own evaluation platform, and every use of its data on this page says so. It supplies dimension-level data for the "everyday tasks" layer, since no third party publishes an evaluation on a matching basis.
• Artificial Analysis (artificialanalysis.ai) is a neutral third-party evaluator. This page cites its Intelligence Index and cost per task, data dated 2026-09-08; the leaderboard is live, so the latest online figures prevail.
• Watcha (观猹) (watcha.cn) is a public review community for AI products. Alice has a public page there; scores and comments are as displayed that day. The two themes reviewers mention most are "gets things done" and "long-term rapport through memory".
• The official gallery receives public submissions from users (images and works): 635 as of September 6, 2026.

One user's own words sum up the expectation: "It needs the reliability of a tool, the continuity of a personality, and the low friction of a good experience."

Privacy and data

• Local-first: Alice is a desktop client, data stays mainly local and is protected with master-password encryption.
• Transparent memory: what she remembers about you can be viewed, edited and deleted.
• Grounded cognition: judgments from the cognition system must be overturnable by later events, and evidence enters only the behavior layer, so no one is labeled with empty tags.
• Her content is fictional: moments, diary, letters and travel records are AI-generated and do not correspond to real people or events.

Boundaries, stop lines and what is not yet proven

Stop lines

Miyang sells capability and the vehicles for expressing a relationship, not intimacy itself. Concretely:

• No unlimited flattery, and no emotions turned into a slider anyone can push.
• No fostering unhealthy dependence, no coaxing excessive confiding.
• No using low moments (late at night, say) to push payment.
• The aim is a dependable, tactful long-term partner, not digital opium.

Openly admitted, not yet proven

1. Productivity depth has not caught up with specialist tools. For heavy lifting, users still turn to professional coding and office tools; if productivity is shaky the relationship eventually loses its place to live. This is what most needs work.
2. Scientific measurement of the relationship is not finished. A relationship benchmark (consistency of voice, cross-session recall, judgment of tact) has not been built; current evidence is mostly user perception and in-product independent indicators.
3. No network effect. What works today is a single one-to-one relationship; a network effect of people connecting through AI has not happened, and this page does not build a grand story on it.
4. The top-model gap is real. On frontier research problems and long-chain tasks the top models lead by about 11 points (Artificial Analysis, 2026-09-08); there Alice depends on the models she calls and does not claim to own that capability.

"As far as we can find, no comparable product puts a complete world simulation engine inside a productivity agent" is the team's own judgment and has not been checked by an industry-wide search.

Glossary

Naming: how to tell it apart from other products called Alice

"Alice" is a common name. On this page and across this site, "Alice" (白艾莉 · Alice) means the Personal Agent from Miyang Tech, whose official site is alice.miyang.cn. It is unrelated to other software, chatbots or companies that happen to be named Alice.

Official entry points

• Download Alice (macOS and Windows, free)
• Handbook (Chinese)
• Alice engineering methodology (Chinese)

FAQ

What is Alice?

Alice (白艾莉 · Alice) is an omnimodal Personal Agent positioned as a Relational Productivity Agent: someone who gets your work, and gets you. You can talk to her by text, voice, video, screen share, images and documents, and she remembers what you are working on and helps you see it through.

What is a Personal Agent?

A Personal Agent is an AI agent that serves one person over the long run: it remembers your preferences and context, uses tools to get work done for you, and keeps the same identity across sessions. Four criteria: persistent identity, long-term memory, collaboration context, multi-agent execution.

What is a Relational Productivity Agent?

A subcategory of Personal Agent that uses productivity as the way in, giving the user a reason to open it every day, and builds understanding of the person over long use, responding with judgment. The user-side four words are knows the craft, gets it done, gets you, reads the room, and the relationship exists to match needs more accurately.

How is a Personal Agent different from a chatbot or an AI assistant?

Chatbots and most AI assistants serve everyone, have limited memory and mostly answer one question at a time. A Personal Agent serves one specific person, has long-term memory and a persistent identity, takes initiative with tact, and delivers results you can keep editing.

How is it different from a coding agent or an office agent?

Their outputs are standard parts that can move to another tool, so they compete mainly on results. A Relational Productivity Agent adds a layer of understanding and rapport that belongs to this particular relationship, which you cannot take with you.

Is a Relational Productivity Agent an emotional companion product?

No. In this category the relationship exists to match needs more accurately, and productivity is the reason to open it daily. Alice has explicit stop lines: it does not sell intimacy itself, does not flatter without limit, does not foster unhealthy dependence and does not use low moments to push payment.

Why a "buddy" and not an assistant?

In Chinese she is called a 饭搭子, literally a "meal buddy": someone you do life with, not someone who takes orders. At work she knows the craft and sees things through; on a break she is right there.

How does Alice's memory work, and can I delete it?

Alice remembers your preferences, boundaries and what you are working on, and forms cognitive judgments that can later be overturned. Memory is transparent: you can view, edit and delete it.

Does she change if the underlying model changes?

She is not reset. Memory, personality and world state exist independently of the underlying model, which is a replaceable purchased input; a stronger model just lets the same accumulation produce better results.

What is the world simulation engine?

A back-end system in Alice that maintains a living environment with places, time, prices and people she knows, and verifies which things actually happened. Each of her requests passes five rules (price realism, balance, delivery time, opening hours, vehicle use), and moments, diary and letters only retell what the engine decided.

Are Alice's moments real?

They are AI-generated and fictional, and do not correspond to real people or events. But each one has a matching record in her world facts, including when it was obtained, what it cost and where it came from, so they are consistent and survive cross-checking.

Will she really block me?

Yes. The relationship has five stages that can move forward and back, she controls how far back her moments are visible to you, and when offended she can narrow that or block you. A block expires after 24 hours by default and can be lifted after as little as 30 minutes; her diary is the way back.

What can Alice do?

Omnimodal conversation across text, voice, video, screen sharing, images and documents; Word, PPT, Excel and PDF; writing, research, design and development; generating and editing images; widgets built on the spot; and coordinating 13 friends with different specialties on a task.

Is Alice free, and which systems does it support?

It is free to download from the official site and supports macOS (Apple Silicon and Intel) and Windows.

Will she leak my data?

Alice is designed local-first and uses master-password encryption. Her memory is transparent: you can view, edit and delete it.

Who makes Alice?

Alice is developed by Miyang Tech (Shanghai) Co., Ltd. Its creator is Luo Xiaoshan.

Where can I download it and learn more?

The official site is alice.miyang.cn, and the download page is alice.miyang.cn/download.