AI Companions Get Persistent Memory
New AI developments are pushing the technology into more personal realms, with the emergence of real-time "AI waifus" or digital companions. A key breakthrough is persistent memory, which allows the AI to retain information across conversations for richer, more personalized interactions. The same report highlights the debut of Qwen 3.5, a next-gen model with enhanced reasoning and lower latency.
The global AI companion market was valued at $37.12 billion in 2025 and is forecast to grow to $552.49 billion by 2035. This expansion is driven by factors including rising loneliness and increased demand for mental health support, with young adults currently representing the largest user demographic. Persistent memory functions by storing information externally, often in a vector database, rather than within the AI's core model. This allows the AI to move beyond short-term, session-based "episodic memory" and recall user preferences, facts, and conversation histories across multiple interactions over long periods. The "waifu" phenomenon originates from the 2002 anime *Azumanga Daioh* and describes a fictional character for whom a person has deep, often romantic, affection. AI companions tap into this by providing a non-judgmental space for emotional support and connection, addressing human needs for intimacy without the complexities of real-world relationships. The Qwen 3.5 model is developed by Alibaba Cloud's Qwen team. It is an open-source model with 397 billion parameters, but it uses a sparse Mixture-of-Experts (MoE) architecture that only activates approximately 17 billion parameters per token, significantly reducing computational cost and latency. This efficient MoE architecture is combined with Gated Delta Networks to enhance reasoning speed. This allows the model to perform more complex, multi-step thinking, similar to a human's "chain of thought," without the long delays that typically make such processes impractical for real-time conversation. Beyond text, Qwen 3.5 possesses native vision-language capabilities. This enables it to understand and interact with graphical user interfaces, meaning it can see a screen, identify clickable elements, and control desktop or mobile applications to complete tasks for a user.
