Liquid AI Bets on Device Context for Personal AI Agents

Most AI models today are built on the assumption that more compute is always available. Need more capacity? The cloud scales up. Liquid AI Inc. is working from the opposite assumption: personal AI belongs on the hardware people already carry, and that hardware has hard limits.

Jeffrey Li, the company's chief operating officer, laid out that view in a conversation with theCUBE Research's Dave Vellante and John Furrier at the Fully Connected event. TheCUBE is the livestreaming studio of SiliconANGLE Media. The discussion covered on-device AI agents, fixed edge compute and what Li calls observability loops.

The case for the edge

Li's argument is simple. The edge gives you less compute, but it gives you a much better picture of the user. That trade, in his view, makes devices the natural home for personal AI.

"The vision we have is that we should bring AI ... closer to the user," Li said. "So we focus on building these AIs to run on the devices all around us."

He was specific about which devices he means. "What is the best form factor to capture that signal? It's the devices in our pockets," he said. "It's our phones, it's our wearables, it's our watches, it's our PCs, it's our cars."

What Liquid Context does

The company's product here is Liquid Context, which is optimized for Snapdragon processors. It sits in the middle of the stack, between models, agents and hardware. According to Li, it uses signals from the device to work out who the user is and what they are trying to get done.

That matters because of how agents are built. The software that turns a model into a working agent - often called a harness - also has to keep track of user context. A common approach is to store that context in a text file that keeps growing. On a phone or a watch, Li said, that approach runs into trouble quickly.

"The problem with devices is that you have fixed compute," he said. "You have to fit within the zero-sum compute. That means a lot of the assumptions around how harnesses today are built no longer hold at the edge."

Liquid AI's answer is to use its own models to decide what information is worth keeping and how to compress it, rather than letting the context pile up.

From deployment to self-improvement

The company is not only talking about phones. Liquid AI is working with Mercedes-Benz Group AG to bring on-device AI into its cars.

Li said the next challenge is keeping agents in line with what users expect long after they ship. "From here, we want to build these systems and these agents to be able to self-heal and improve and personalize on their own over time," he said. "We're building observability loops and continuous improvement loops that will improve both the model and the harness over time through natural usage."

In practice, that means two things are meant to get better with use: the model itself and the harness around it.

Our Take

Liquid AI's pitch adds to a steady push toward local AI. Google's offline meeting tool for Mac and new hardware like a smart ring designed for AI agents point in a similar direction. The interesting part of Li's argument is not the hardware, though. It is the claim that agent harnesses built for the cloud do not carry over cleanly to devices with fixed compute.

That suggests memory management could become a key differentiator for on-device agents. The self-improvement goal is still described as work in progress. It is worth watching how "improve through natural usage" is handled in practice, especially around what data stays on the device, and whether the Mercedes-Benz work produces concrete products.