UniCone: A Proactive AI Partner for Creators

Inspiration

AI models have become remarkably capable, yet most people still use them through short, isolated conversations. Users repeatedly explain their background, copy information between tools, and decide every time when and how AI should help.

This creates a gap between what AI can technically do and the value it delivers in everyday work. The problem is especially visible among creators, consultants, lawyers, entrepreneurs, and other independent professionals who manage multiple projects, produce content, and rely on personal relationships to create opportunities.

We created UniCone to close this gap. Our goal is to build an AI partner that continuously understands the user’s context and provides useful assistance at the right moment.

What it does

UniCone is a proactive AI agent designed for creators and other “super individuals”—people who generate value through their expertise, content, reputation, and professional networks.

With the user’s permission, UniCone connects context across devices and modalities, including screen activity, applications, files, browser information, conversations, location signals, and calendar events. It transforms this information into structured events and long-term memory.

Instead of waiting for a detailed prompt, UniCone can proactively:

• Identify relevant trends and content opportunities
• Recommend personalized topics based on the user’s real experiences
• Generate outlines and ready-to-record scripts
• Organize meetings, shoots, deadlines, and publishing schedules
• Prepare meeting briefs and reminders
• Draft messages, emails, and other materials for user approval
• Help users retrieve relevant people and opportunities from their networks

The user remains in control of final actions. UniCone prepares and recommends, while the user decides what to approve, publish, or send.

How we built it

UniCone uses a three-stage architecture: sensing, intent understanding, and service delivery.

The sensing layer collects authorized multimodal context from desktop and mobile devices. Visual information can be interpreted using vision-language models and OCR, while audio, application activity, browser context, files, calendars, and location signals provide additional context.

The memory system converts raw sensing streams into structured event streams. It removes redundant information, groups activity into meaningful events, and cross-checks multiple data sources to reduce errors. These events are distilled into a hierarchical user profile representing current priorities, habits, workflows, preferences, knowledge, decisions, and professional relationships.

The proactive-service layer combines this internal context with relevant external information. A lightweight model first determines whether assistance may be useful, and a larger language model analyzes the situation and prepares an appropriate response. This two-stage system reduces unnecessary interruptions and controls computational cost.

Our desktop product has been released as an open-source project and has received more than 2,400 GitHub stars. We are also developing UniCone Pod, an always-on wearable sensing device, and UniCone Base, a local computing hub for users who require private on-device processing.

Challenges we faced

One of the hardest challenges is determining when AI should intervene. If the agent speaks too often, it becomes distracting. If it reacts too slowly, it misses the opportunity to help. Our research therefore focuses on both the timing and content of proactive assistance.

Long-term memory is another challenge. Personal context is multimodal, continuous, and constantly changing. We address this through semantic event segmentation, information compression, cross-source verification, hierarchical memory, and graph-based indexing.

Privacy is equally important because personal context may contain sensitive professional and personal information. UniCone uses explicit authorization and gives users control over sensing and final actions. We are also developing local computing hardware for users who require private deployment.

Mobile operating systems impose additional limits on continuous sensing. UniCone Pod helps capture authorized offline audio context that phones cannot reliably collect, while low-power voice activity detection helps extend battery life.

What we learned

We learned that context can be more important than the underlying foundation model. Even a powerful model produces generic results when it does not understand the user’s history, goals, workflow, or personal voice.

We also learned that proactive assistance should support human judgment instead of replacing it. Users want AI to identify opportunities, organize information, and prepare high-quality drafts, while keeping consequential decisions and final actions under human control.

Finally, we learned that building deeply for one user group creates stronger value than offering shallow features to everyone. We are therefore starting with content creators and independent professionals, where our team has direct experience and access to strong user communities.

Progress and impact

UniCone has opened 390 beta-user places and is working with an initial group of 100 creators and independent professionals whose combined audiences exceed 100 million followers.

The product already has paying users. Our private-deployment and enterprise-agent projects have generated more than RMB 1 million in collected revenue, covering the team’s current operating costs.

Our team combines AI researchers from Tsinghua University with experienced content creators who have more than three million followers across platforms. This combination allows us to develop the underlying technology while testing it in real creator workflows.

Our research includes ProAgentBench, a benchmark for evaluating proactive agents with real-world data, and FreeU, a framework for balancing awareness and autonomy in proactive assistance.

What is next

Over the next six months, we will focus on three connected workflows for creators: content production, schedule orchestration, and professional-resource matching.

We will improve the sensing and memory architecture, expand real-world user testing, launch the mobile product, and continue developing UniCone Pod and UniCone Base.

Our long-term vision is to give independent professionals a trusted AI partner that understands their context, helps them make better use of their knowledge and relationships, and proactively supports their work without taking control away from them.

Built With

• continual-learning
• deepseek
• desktop-applications
• knowledge-graphs
• large-language-models
• long-term-memory
• mobile-applications
• multi-agent-systems
• multimodal-ai
• ocr
• on-device-ai
• private-ai-deployment
• proactive-ai-agents
• reinforcement-learning
• semantic-event-segmentation
• speech-processing
• vision-language-models
• voice-activity-detection
• wearable