Creating an AI like Candy AI is a challenge in itself. Although the idea is quite simple, creating a candy ai clone that is emotionally responsive and scalable is not that easy.

The first challenge that comes with AI development is the management of contextual memory. The AI needs to remember the previous conversations, emotional expressions, and user preferences without being repetitive or inconsistent.

Another challenge that comes with AI development is personality consistency. The AI companions need to behave in a predictable yet dynamic way. The AI needs to be consistent in its behavior and tone without being out of character or tone.

Another challenge that comes with AI development is content moderation and safety. The developers need to ensure that the conversations are not only expressive but also within the boundaries of ethics and laws. This becomes even more challenging when the conversations become more personal and emotional.

From an engineering perspective, the issue of scalability in infrastructure is always present. The need for real-time AI answers, the number of concurrent users, and the logic of personalization requires robust backend infrastructure and effective use of AI models.

Development teams at companies such as Suffescom Solutions have observed that the key to success lies not in borrowing superficial elements but in addressing the underlying issues. Otherwise, even the best-designed candy ai clone might find it difficult to attain reliability.

If i am not wrong you can try to create a Candi AI system (i never used, but i suppose it similar to character tavern but with chat like approach instead of narrative?) adressing several things. 1) A chat like interface, not like chatGPT but something like messenger/whattsapp or things like that. 

2) have the bot promptet to reply like in a conversation. Things like: 'Hi, how was your day?'  instead of: *I enter the kitchen and saw you there, drinking a coffee as the steam caress your face and the sun paint glowing lines in the background.* "Hi, how was your day?" *bla bla bla* 

3) A Diffusion model (like SDXL or SD1.5) with a character lora and well configured previously to create images. For example if you request the character some n****

4) Perhaps you may also need a reasoning model (Deepseek or Qwen models) even small parameters models tend to give good replies. 

I don't know, those are just a few ideas. I made a prototype for a system that allows LLMs to control NPC's in a 3D game. It's called 'Coppersons' Is not the same as Candy.AI   but maybe it can give you an idea/approach/whatever? it can be used in both 2D and 3D applications.