Many activities in the cybersecurity realm can be represented using graphs stream, such as call graphs. In this paper, we introduce an innovative method to detect Advanced Persistent Threats (APTs) from their onset. Unique to our approach is the ability to assimilate both structural and temporal aspects, crucial for differentiating between benign and malicious activities. To overcome challenges presented by streaming data processing, we leverage hashing techniques for a compact data representation. This method, when combined with a dynamic machine learning framework, facilitates swift, incremental detection and ensures minimal memory usage. Empirical evaluations underscore the efficacy of our approach, allowing a real-time response by pinpointing APTs at the initial stages of their activity.