LLMSense is a system that leverages Large Language Models (LLMs) for high-level reasoning over spatiotemporal sensor traces. Traditional machine learning approaches for high-level reasoning tasks struggle due to limited training samples and high-dimensional sensor data, necessitating integration of human knowledge. LLMSense addresses this by using LLMs' reasoning capabilities and world knowledge to recognize complex events from long-term sensor data. The system includes an effective prompting framework for high-level reasoning tasks, which can handle raw sensor data and low-level perception results. Two strategies are designed to enhance performance with long sensor traces: summarization before reasoning and selective inclusion of historical traces. The framework is implemented in an edge-cloud setup, with small LLMs on the edge for data summarization and high-level reasoning on the cloud for privacy preservation. Results show that LLMSense achieves over 80% accuracy on tasks such as dementia diagnosis and occupancy tracking. The paper provides insights into leveraging LLMs for high-level reasoning on sensor traces and highlights future research directions. LLMSense demonstrates the potential of LLMs in interpreting complex sensor traces and making inferences based on long-term data, offering a trade-off between accuracy, latency, and data privacy. The system is evaluated on two high-level reasoning tasks, showing its effectiveness in handling long sensor traces and improving performance through summarization and selective inclusion of historical data. The paper also discusses challenges in applying LLMs to long-term sensor traces, including converting sensor data into natural language and improving performance with long series of sensor data. LLMSense provides a framework for high-level reasoning on sensor traces, demonstrating the potential of LLMs in this domain.LLMSense is a system that leverages Large Language Models (LLMs) for high-level reasoning over spatiotemporal sensor traces. Traditional machine learning approaches for high-level reasoning tasks struggle due to limited training samples and high-dimensional sensor data, necessitating integration of human knowledge. LLMSense addresses this by using LLMs' reasoning capabilities and world knowledge to recognize complex events from long-term sensor data. The system includes an effective prompting framework for high-level reasoning tasks, which can handle raw sensor data and low-level perception results. Two strategies are designed to enhance performance with long sensor traces: summarization before reasoning and selective inclusion of historical traces. The framework is implemented in an edge-cloud setup, with small LLMs on the edge for data summarization and high-level reasoning on the cloud for privacy preservation. Results show that LLMSense achieves over 80% accuracy on tasks such as dementia diagnosis and occupancy tracking. The paper provides insights into leveraging LLMs for high-level reasoning on sensor traces and highlights future research directions. LLMSense demonstrates the potential of LLMs in interpreting complex sensor traces and making inferences based on long-term data, offering a trade-off between accuracy, latency, and data privacy. The system is evaluated on two high-level reasoning tasks, showing its effectiveness in handling long sensor traces and improving performance through summarization and selective inclusion of historical data. The paper also discusses challenges in applying LLMs to long-term sensor traces, including converting sensor data into natural language and improving performance with long series of sensor data. LLMSense provides a framework for high-level reasoning on sensor traces, demonstrating the potential of LLMs in this domain.