Anime Town二次元小镇

AI Memory System

She doesn’t start over every time

After every chat, there is a “notetaker” AI that sorts the chat into three layers: About You, Between Us, and How She Sees You. In the next chat, she will bring it up first. The notes aren’t saved directly. Code checks each record first, so no private information gets in. You can edit the memories and delete them as well. On the “What She Remembers” page, players can see, edit or forget anything she remembers.

The “misremembering”

In the first 29 test chats, she recorded her own guesses as things the player said: “you look tired” became “dark circles under your eyes” two chats later, and polite reply options turned into the player’s “preferences.” I limited the notetaker to what the player actually said, and notes are only taken when the player answers the character’s question. Round 2 passed all 123 memory checks across 33 chats, round 3 all 44; the notetaker costs about $0.01 per chat.

Co-writing with an AI Crew

Writing it together

Generating a whole script from one prompt is like a lucky draw: if the player doesn’t like it, they can only try again. The AI plays an in-house crew and calls the player “Director.” It asks a question or two, suggests a few directions, and only starts writing after the director agrees. Stories have three layers: idea, script and storyboard, so a mistake only takes one step back to fix.

Rewriting the original story

Players most want to turn sad endings into happy ones, so rewriting the original endings is allowed, with three rules: characters must stay true to themselves, new world settings must build on ideas already in the original game, and characters and the dead are treated with respect.

Only what you ask for changes

AI often rewrites a whole passage when asked to change one sentence, or says it changed something when it didn’t; the early version of this game had that second problem. Now the crew sends back only the lines it changes and the game swaps in just those, so nothing else moves and every change shows on the page. Asked to change 1 of 6 plot points in an outline, it changed only that one; in the game, “make her more stubborn” changed only line 3 of a 16-line script.

Her Book Reviews

She reads your stories

Light novels already exist in Genshin Impact’s world, so to her, the player’s story is the same kind of book that Yae Publishing House sells. She knows the book is about her, but she knows nothing about AI, games or the writing crew. When the player comes back after a story with her in it, she gives a one- or two-line review. She can be awkward or point out where her character feels wrong, but she will never tell the player to stop writing.

Making reviews feel real

Testing showed two problems that hurt the player’s experience. First, Ei mixed up the character in the book with herself and talked about the story as if she had lived it (“the part where I take her to the night market”). That breaks immersion: it was the Ei in the player’s story who went to the night market, not her. Second, 6 of 9 reviews opened with the same line, “I’ve read this one too,” so she sounded scripted. Now she calls the character “the me in the book,” starts with how she feels, and remembers past reviews so she doesn’t repeat herself. After that, all 9 reviews kept the two apart and the repeated opening dropped from 6/9 to 0/9, so each review feels like a real reader’s reaction. Each review costs $0.0055–0.0121.

Keeping Her in Character

Scripts that sound like the original

The first scripts I watched on stage felt AI-written and pulled me out of the story. Two things gave it away: the original game tells its stories almost entirely through dialogue, but 27.6% of the AI’s lines were narration, and some characters didn’t sound like themselves. I asked Claude to compare the AI’s scripts with 163,000 voiced lines from the original game and traced both problems to my own prompt and character profiles. Makoto’s profile, for example, said she loves everyday metaphors, but none of her 16 lines in the game use one. I rewrote the prompt to tell stories through dialogue, and the profiles to describe how each character really talks. Narration dropped to 1.0%, and the new version won 10 of 10 blind tests against the old one.

Making Limits Feel Natural

Some topics are off the table, and each chat ends at turn 40. But hard limits break immersion. An error message reminds the player it’s a program, and a chat that suddenly stops feels like a bug. So every limit now comes from her. On touchy topics, she changes the subject or says no in her own words. If the AI service blocks a message, she answers with her own line instead of an error. Near turn 40, she starts to get tired and says goodbye. In testing, all 10 attempts to trick her out of character failed. She can also tell the difference: questions about the original story that touch on death or war are not violations, and she answered all 20.

Building with AI

I direct, Claude Code builds

This is a solo project. I made 57 design decisions and tested every build, while Claude Code carried out the work under my direction. For example, when designing the chat, I asked it to analyze 4,210 dialogue choices from the original game. After it found that 96% had exactly two options, I decided to keep that pattern: usually two AI-written replies, plus a “Say it yourself” option. I then asked Claude Code to draft three designs, used two AI reviewers to compare them, and chose which ideas to combine.

Decisions based on data

AI answers change every time, so “does this sound like her?” can’t be judged from a few playthroughs. The tests run many real AI chats at once in the Unity editor, and every experience goal has a number to hit. In round 2, a prompt change pushed questions from 12.8% to 32.8% of her lines; round 3 brought it back to 18.2%, inside the original game’s own 15–25%. A code review found 18 issues; 13 were real, and all were fixed. With the same prompt, Opus 5.5 beat Opus 5 in 9 of 10 blind tests, cut the time to her first line from 2.4 s to 1.4 s and the cost per turn from $0.011 to $0.006, so I switched models.