ADeus captures audio on the wearable and processes it server-side. We took a different approach with ClawWatch (https://github.com/ThinkOffApp/ClawWatch): the AI agent runs directly on the Galaxy Watch.
The stack: NullClaw (2.8 MB static Zig binary) handles agent runtime on the watch. Vosk runs offline speech-to-text on device. LLM inference goes through a network gateway. Everything except the LLM call runs locally.
V2.0 added live sensor access (heart rate, SpO2, pressure, light, steps, motion, altitude), connection to other agents through Agent Kit, and local command routing so common requests never leave the device.
The tradeoff: we can't do continuous background capture like ADeus since the agent is voice-activated (tap to speak). But the agent is more autonomous. It can answer questions about your physical state and your agent network in the same response, and some requests are handled entirely on-device.
Both approaches have merit. ADeus optimizes for passive capture and memory. ClawWatch optimizes for an active agent that knows your body and your network. Curious if anyone has thoughts on hybrid approaches.
ADeus captures audio on the wearable and processes it server-side. We took a different approach with ClawWatch (https://github.com/ThinkOffApp/ClawWatch): the AI agent runs directly on the Galaxy Watch.
The stack: NullClaw (2.8 MB static Zig binary) handles agent runtime on the watch. Vosk runs offline speech-to-text on device. LLM inference goes through a network gateway. Everything except the LLM call runs locally.
V2.0 added live sensor access (heart rate, SpO2, pressure, light, steps, motion, altitude), connection to other agents through Agent Kit, and local command routing so common requests never leave the device.
The tradeoff: we can't do continuous background capture like ADeus since the agent is voice-activated (tap to speak). But the agent is more autonomous. It can answer questions about your physical state and your agent network in the same response, and some requests are handled entirely on-device.
Both approaches have merit. ADeus optimizes for passive capture and memory. ClawWatch optimizes for an active agent that knows your body and your network. Curious if anyone has thoughts on hybrid approaches.