The LLM driven IoT applications need to handle time-sensitive information and dynamically evolving data. For example, the sensors in IoT systems send real-time data that must be taken into consideration by the LLM-driven application. The traditional architectures store such real-time IoT data in a database and re-prompt the LLM by suitably in- corporating the new information on each request (or at designated time). Such an approach becomes inefficient especially when: (1) Data arrives at a high rate, e.g., streaming updates, large volumes of logs. (2) Only some fraction of updates are relevant to the user’s current question or context. (3) The application must preserve the conversation flow in a manageable size while still allowing new data to “flow in.” In this paper we present an architecture that addresses these challenges by maintaining a lightweight “timeline” of recent and relevant “events”. This design ensures that the LLM always sees a relevant snapshot of the most critical information at each step, without needing to ingest the entire stream of events every time. An important element in our architecture is the “event relevance scoring” module which allowed us to maintain a “sliding window” of the most crucial recent incoming sensors data. Our experiments show that, compared to the baseline, the proposed approach’s response time is 15.41% faster and it generates 65.16% fewer tokens, which reduces token-based costs significantly. On average, it uses 5.41% less GPU and consumes 5.13% less CPU.