Liquid AI builds personal AI with device-level context awareness

What changed

Liquid AI announced its approach to building on-device personal AI context awareness systems that operate within the hardware constraints of edge devices like phones, wearables, watches, PCs, and cars. Jeffrey Li, chief operating officer of Liquid AI, discussed this work at the Fully Connected event, explaining how Liquid Context software sits between AI models, agents, and hardware to enable personal AI deployment without cloud dependency.

The company released Liquid Context, which is optimized for Snapdragon processors. According to Li, the software uses device signals to build an understanding of who the user is and what they are trying to accomplish. The approach acknowledges a fundamental shift: personal AI belongs on edge devices because they offer a richer view of the user than cloud-based systems, even though they operate under fixed hardware constraints.

Li stated that the edge offers a far richer view of the user compared to cloud systems. Devices in users’ pockets, phones, wearables, watches, PCs, and cars, capture this signal directly. This makes them the natural home for personal AI, according to Li’s vision of bringing AI closer to the user.

Why it matters

The significance of on-device personal AI context awareness lies in how it changes where user context should reside. Professionals deploying AI tools face a practical question: should user context live on edge devices or in the cloud? Liquid AI argues that edge devices should be the natural home for personal AI because they capture device signals directly, from location to application usage patterns to communication history.

This affects several groups. Device manufacturers need software that allows them to run AI agents without offloading data to the cloud. Enterprise deployments benefit from keeping sensitive user context local. End users potentially gain AI systems that improve over time without sending personal information to remote servers.

The fixed compute constraint is real and fundamental. On-device systems cannot expand resources elastically the way cloud systems can. Liquid AI addresses this by using its own models to decide which information to retain and how to compress context. This differs from cloud-based AI agents, which can maintain growing context windows because they have virtually unlimited storage and compute available.

Agent harnesses, the software that turns models into functioning agents, manage user context. Keeping that context in an ever-growing text file poses challenges on devices with limited resources. Liquid AI’s approach is to compress context intelligently rather than store everything.

The company is collaborating with Mercedes-Benz Group AG to bring on-device AI to vehicles. This partnership illustrates a practical application: cars need personal AI agents that understand driver preferences and behavior without transmitting that data off the vehicle. Li noted that keeping agents aligned with user expectations long after deployment is the next step.

What to test

Before adopting Liquid AI’s approach or similar on-device personal AI context awareness systems, professionals should verify several vendor claims through testing:

Compression trade-offs. Liquid AI says its models decide which information to retain and compress context to fit fixed compute limits. Test whether the system actually retains the information that matters for your use case. Verify that compression does not degrade agent performance over time as context accumulates.

Improvement without retraining. The company claims it is building observability loops and continuous improvement loops that will improve both models and harness software over time through natural usage. This is a forward-looking claim. Test whether agents actually perform better after extended deployment, and measure whether improvements happen automatically or require intervention.

Hardware fit. Liquid Context is optimized for Snapdragon processors. If your deployment uses different hardware, test whether the software performs as described on your target devices. Fixed compute constraints vary significantly across device types.

Context window limits. Ask for specifics about how much context the system can maintain on your target hardware. Compare this to the context window sizes of cloud-based AI agents you might otherwise use.

Integration with existing tools. Agent harnesses turn models into functioning agents. Test how well Liquid AI’s harnesses integrate with your existing development tools and deployment infrastructure. Li indicated that assumptions about how harnesses are built at the edge differ from cloud assumptions.

Self-healing behavior. Li stated that the company wants agents to self-heal and improve and personalize over time. Test whether this actually occurs in production environments or if it requires manual tuning.

The conclusion

Liquid AI is addressing a real engineering problem: how to run personal AI agents on edge devices within fixed hardware constraints. The company’s focus on device-level context awareness and compression represents a practical approach to on-device AI deployment.

The approach makes sense for use cases where user privacy or offline capability matter. Vehicles are a clear example, as they need to function locally and should not transmit driver behavior data to external servers. Phones and wearables present similar requirements.

What to monitor next: whether Liquid AI’s continuous improvement loops actually work in production. The company says agents will improve over time through natural usage, and harnesses will improve alongside models, but this remains a forward-looking claim. Real-world deployments will show whether these systems actually self-heal and personalize as described. The Mercedes-Benz collaboration will serve as a practical test of these capabilities.

Professionals considering on-device personal AI should also track how the company evolves its compression algorithms. As agents run longer and accumulate more context, maintaining performance consistency on fixed hardware becomes more difficult. The technical details of how Liquid AI solves this problem will matter more as these systems reach scale and encounter longer-running deployments.

The fixed compute constraint Li identified represents a genuine architectural difference from cloud-based systems. Understanding how Liquid AI’s approach performs when this constraint becomes tighter will be essential for professionals evaluating whether on-device personal AI is ready for their use cases.