Terminus Group and BUTTONS Join Forces to Launch Groundbreaking Consumer AI Agent

Terminus Group has developed a groundbreaking intelligence system, HALI, a next-generation AI agent system that marks a significant leap in consumer-facing artificial intelligence. By pioneering the integration of end-to-end reinforcement learning with curated industry datasets, HALI achieves tool invocation, sub-task decomposition, and multi-agent collaboration without human intervention, significantly enhancing autonomy, flexibility, and generalization.

HALI's first commercial implementation arrives via collaboration with BUTTONS, the global intelligent luxury brand. Integrated into the newly released BUTTONS CLIP audio wearable, the system delivers targeted AI functionality, enabling personalized intelligent interaction, serving as a proof-of-concept for future AI-driven lifestyle enhancements. Moving forward, Terminus Group and BUTTONS will jointly expand the HALI ecosystem into a comprehensive suite of intelligent hardware, from AI glasses and smart rings to smart speakers and home robotics.

Last year, on the center stage of Web Summit Lisbon 2024, Terminus Group shared its vision for the four essential capabilities of general-purpose AI Agents: human-like reasoning, long-term memory, multi-agent collaboration, and high-dimensional perception - principles that guide HALI’s design.

1. Human-like Reasoning - Eliminating Manual Workflow Dependencies

Most current AI agents rely on predefined workflows, splitting tasks into subtasks, processing them separately, then compiling results. While this approach can be effective, it limits adaptability and scalability due to its reliance on human orchestration.

Terminus Group addresses this challenge by leveraging end-to-end reinforcement learning (RL), training HALI with high-quality data to develop adaptive strategies through reward functions rather than manual programming. This enables the agent to reason, plan, and execute tasks directly from user input, without relying on predefined workflows or tool-calling logic, unlocking greater autonomy and flexibility in real-world interactions.

With this approach, HALI achieves both elegant simplicity and robustness, significantly improving its generalization and accuracy. While most existing agent products focus on engineering improvements, Terminus Group sees vast potential in advancing the model-level capabilities. Reinforcement learning marks an early step in building agents with human-like cognitive abilities.

2. Long-Term Memory – Balancing Cost and Efficiency

Long-term memory is a core capability of AI agents, especially for those requiring real-time, low-latency interactions. Traditional methods store raw conversation history, leading to bloated data, slower retrieval, and higher costs over time. Moreover, over time, storage costs skyrocket, and long context windows slow down inference, impacting the performance and user experience of the agent products.

HALI leverages an advanced data compression strategy to retain only the most critical information, filtering out unnecessary noise to optimize performance. Paired with knowledge graphs that greatly enhance intent recognition and retrieval accuracy, this dual system dramatically reduces storage and computing costs while significantly improving speed and relevance. Together, these innovations make long-term memory not only possible but practical for next-generation, companion agents.

3. Multi-Agent Collaboration – Optimizing Real-Time Performance

Seamless coordination between tools and models is vital for user experience. Terminus Group adopts a hybrid architecture using multiple models in parallel (on-device light models and server-based large models). This setup ensures both low-latency responsiveness and high accuracy.

The light models, deployed on-device, are trained to understand basic control commands using accumulated user data, allowing direct local execution and avoiding network delays. For complex tasks, powerful server-based large models are invoked to ensure accurate understanding and execution.

Partnering with Global Intelligent Luxury Brand BUTTONS to Revolutionize Intelligent Luxury Experiences

The HALI system is now available on mobile and PC. The first HALI-powered intelligent luxury earbuds, BUTTONS CLIP, has been officially launched, offering users a taste of this next-gen AI experience.

The BUTTONS CLIP represents a paradigm shift in wearable technology, blending aerospace nitinol and hypoallergenic silicone to achieve both luxury aesthetics and all-day comfort. While the industry remains fixated on technical specifications like sound quality and decibel reduction, BUTTONS has redefined the very concept of personal audio by creating what it calls an "emotional mobile music lounge", a space where technology adapts to human emotions.

The HALI intelligent system brings this vision to life through seamless user interaction. Following a quick one-time setup, users can instantly activate their AI companion with a simple "HALI HALI" voice command. Unlike conventional voice assistants that rely on rigid programming and manual updates, HALI functions as a truly autonomous system, continuously learning and evolving with each interaction. This technology transcends conventional smart devices, creating what we term "cognitive wearables" - a new category that fulfills luxury consumers' desire for meaningful connections with their technology.

Building on this foundation, BUTTONS and Terminus Group will expand their collaboration to develop an ecosystem of HALI-powered devices, including smart glasses, rings, speakers and next-generation home robotics solutions. This expansion comes at a pivotal moment in AI development, where large language models are achieving unprecedented capabilities in knowledge processing. HALI AI agent serves as the crucial bridge between these digital capabilities and physical-world applications.

Terminus Group's development strategy rests on parallel advancement in two critical areas: breakthroughs in foundational model architectures and excellence in systems engineering implementation. This dual-track approach ensures both conceptual innovation and practical deployment capability progress in lockstep. While the current HALI platform demonstrates strong core functionality, its physical-world perception remains an area for continued development. The engineering roadmap focuses on integrating additional sensory modalities, expanding into specialized vertical applications, and cultivating increasingly sophisticated datasets. These efforts will progressively enhance the system's capacity to interpret and navigate the multidimensional complexity of real-world environments, bringing us closer to truly contextual artificial intelligence.