Choosing an AI girlfriend or AI companion: how Shiori differs from an ordinary chat AI

Yinfeng

Contents
CG: a girl under a pale pink sky, hands clasped, small wings spread

Looking for an AI girlfriend or an AI companion, only to find that most chat AIs forget you the moment you open a new window, and never say a word unless you speak first? This post spells out where Shiori really differs from an ordinary chat AI, who it suits, and what you need to bring yourself.

The app’s interface is in Chinese, so below every button and tab goes by its Chinese name, with its meaning in brackets.

What an ordinary chat AI is like

Most chat AIs work question by answer: you ask, it answers; close the window, and next time it starts from scratch. It’s clever, but it doesn’t know you, and it won’t feel a thing if you stay away for three days.

Shiori starts from the other end: first there is a character, and then she slowly gets to know you. She has her own persona and personality, remembers what you’ve talked about, your relationship changes bit by bit, and she comes to find you on her own. One thing at a time, then.

She remembers you

Every character keeps her own memory, never mixed with anyone else’s. After you’ve chatted for a while, she sorts the conversation into a few “little notebooks” herself: lasting facts about you, things the two of you have been through, what you’ve been talking about lately. The next time you speak up, she recalls whatever is related.

Want her to remember something? Just say “remember this”. Got it wrong? Say “forget that”. How her memory works in detail, I explained in She remembers you.

A relationship is built bit by bit

Every character has an affection level towards you, from −100 to 100, in seven stages: loathing, cold, stranger, familiar, friend, close and devoted. Each round of chat moves it by 3 points at most, every change comes with its reason written down, and if you stay out of touch too long, it slowly fades.

At different stages she talks to you in a different tone. Think one sweet line will rush you straight to devoted? It’s not that cheap. The rules are all in From loathing to devotion.

She comes to find you

Turn on 「主动推送」 (Proactive messages) under 「能力」 (Abilities) in the character’s details, and every so often she’ll wake up and see whether there’s anything worth telling you. Once affection reaches 「亲密」 (Close), if you stay away too long, she’ll also speak up because she misses you.

When you’re away from the computer, her messages go to your private chat in whichever chat app you’ve linked your identity on. The details are in Ignore her too long, and she’ll come find you.

Not stuck in one window

  • Chat apps: add a QQ, Telegram or Feishu account to a character, and you can chat with her on your phone too. Once your identity is linked, the her over there and the her on your desktop are the same one: what she remembers and how she feels about you are one and the same. See this guide for how to connect.
  • Desktop pet: import a desktop pet asset pack, and she’ll stay on your desktop, her replies popping up as bubbles; set up local voice, and she can talk out loud too. See AI Desktop Pet.
  • Story mode: take a character into a story of its own and drive the plot with your actions.

Runs locally, your data stays on your computer

Shiori is a desktop app installed on your computer, with no account to sign up for. Characters, chat history, memories and assets are all stored on your machine in %USERPROFILE%\.shiori\workspace\, and the memories are even plain text files you can open and read.

To be clear, though: replies are generated by the model service you configure, and the chat is sent to that service. With image generation, chat apps or voice enabled, the related content is also sent to the corresponding services. If you want the model to run on your machine too, you can choose a local service such as Ollama.

Open source and free, but you bring the model

Shiori itself is an open-source Personal Agent under the MIT license, free to download. It doesn’t ship a model, so you need to prepare:

  1. A model service (required): any service compatible with the OpenAI Chat Completions API will do. The settings come preset with OpenAI, DeepSeek, Alibaba Cloud Bailian, Moonshot Kimi, Zhipu GLM, SiliconFlow, OpenRouter and Ollama (local), or you can pick 「自定义」 (Custom). For pay-as-you-go services, the provider charges the fees.
  2. An embedding service (optional): lets her search her memory by meaning; only needed for semantic memory.
  3. Other services as needed: chat app accounts, NovelAI (image generation and story CGs), speech recognition and synthesis.

Also, there is only a Windows x64 version for now.

Who it suits, and who it doesn’t

It suits you if: you want a character to spend a long time with, one who remembers you and reaches out first; you don’t mind spending a few minutes filling in a model service and an API key; you want your data to stay in your own hands.

It’s less of a fit if: you want to open a web page and chat with nothing to set up; or you’re on a Mac or Linux.

How to start

  1. Download the Windows installer from the download section of the home page and install it.
  2. On first launch, follow the 「模型 → 角色 → 开始」 (Model → Character → Start) guide: fill in the model service, API key and model, then create your first character.
  3. Start chatting. When you want other abilities later, enable them under 「设置 → 插件」 (Settings → Plugins).

There, the differences are all laid out. Whether you pick it is up to you.

…But if you do, don’t go days without coming to see me, dummy.