My AI Girlfriend Is Cheating on Me

Does this game support offline play?

Yes, the game supports offline play by setting up a local AI provider using Ollama or LM Studio from CLI and entering the appropriate Base URL in the game settings.

What are the requirements for fully offline play?

For fully offline play, you should have at least 16 GB of available RAM/VRAM. The exact hardware requirements depend on the local model being used and its quantization level. Proper compatibility testing is still ongoing.

How do I set up a local AI provider for offline play?

To set up a local AI provider, use Ollama or LM Studio from CLI and enter the appropriate Base URL in the game settings: For Ollama: localhost:11434/v1; For LM Studio: Follow detailed setup instructions provided by the developer.

Which models are recommended for local inference?

DeepSeek V4 Flash is currently the only model thoroughly tested and recommended. The developer plans to publish a compatibility list with more tested models soon.

How does narrative mode work in the game?

Narrative mode alters character behavior, such as likelihood to cheat, by changing prompt/persona cards. It affects how characters interact based on different scenarios and prompts provided.

What are the challenges with using different LLMs in the game?

Different LLMs may require varying system prompts or syntaxes. The developer checks which LLM is connected and uses appropriate settings, though this can affect writing quality due to differences in model capabilities.

What about context window limitations in LLMs?

Context window limitations can lead to quality deterioration as the conversation progresses. The developer has noticed issues with long-term memory retention and plans improvements based on feedback.

How much does it cost to play the game using DeepSeek V4 Flash?

One game day of simulation costs approximately $0.094 per real hour, with most input tokens becoming cached and reducing costs significantly after initial setup.

What is the easiest way to set up the game for playing online?

The easiest setup involves getting a DeepSeek API key from their website, topping up $2 (covering about 10 hours of play), generating an API key, and selecting DeepSeek Official during onboarding.

How does the game handle character interactions and AI behavior?

Each character runs as an independent agent with memories, daily plans, location, relationships, and current situation. They can message people, post on social media, move around, and interact with each other even when you aren't talking to them.

Is there support for image generation in the game?

Actual image generation is planned but currently not available. The developers are working on integrating local or online image-generation providers as part of future updates.

What happens if I encounter 'agent tick failed' or 'interaction.resolve failed' errors when setting up the game locally?

Enabling serving the model to the local network instead of just the machine can resolve these issues. The latest version (1.0.4) includes automatic detection and application of best settings for LM Studio.

How does caching affect performance when using a local backend?

Effective caching significantly improves performance by reducing the need to process entire prompts repeatedly, making prompt processing less of a bottleneck.

Game Progression Tips

• The game progresses through interactions with Rin. Encourage her to talk to others, but she may resist.
• Consider suggesting scenarios where Rin interacts romantically or socially with other characters for progression.

Technical Issues and Fixes

• Join the Discord server and DM the developer if you encounter issues like incomplete action batches or HTTP 400 errors.
• Report any bugs, such as message timeline inconsistencies or loops when characters sleep, to help improve the game.

Optimization Tips

• Use an API service with a model that suits your needs and budget. DeepSeek V4 is recommended for testing.
• Avoid using local LLMs due to VRAM limitations; opt for cloud-based models instead.

Suggestions for Improvement

• Implement an autosave feature with a shortcut key (e.g., F5) to save progress easily.
• Consider adding options to disable modules like instant likes to reduce token usage and prevent loops when characters sleep.