Combination interactions require coordinated operations across multiple GUI widgets within a page to complete a functional task, which are common in modern mobile apps. However, due to widget dependencies, dynamic interface transitions, and context-sensitive execution, existing automated GUI testing approaches based on predefined models or state exploration struggle to reliably complete such interactions. LLM-based GUI testing methods improve semantic understanding but often generate inconsistent actions due to the lack of coordination modeling and interaction context. To address these challenges, this paper proposes COMBDroid, an automated GUI testing approach that enables reliable completion of combination interactions through structured planning and dual memory. COMBDroid first detects combination interaction patterns from GUI structure, then generates page-level interaction plans and grounds them into widget-level executable actions. Inspired by Sudoku solving, COMBDroid maintains a dual-memory mechanism, where context memory preserves execution states within a functionality and experience memory captures reusable interaction patterns. COMBDroid is designed as an integrable module that activates when combination interactions are detected and allows existing testing tools to resume exploration after completion. We evaluate COMBDroid on 100 combination interaction functionalities collected from 100 mobile apps, comparing it with 15 state-of-the-art GUI testing baselines. COMBDroid achieves 103% higher passing rate than the best baseline and improves activity coverage by 65%-100% when integrated with existing testing tools. Furthermore, GUI testing tools integrated with COMBDroid detect 37 new crash bugs in real-world apps, with 26 fixed and the remaining confirmed by developers, demonstrating its effectiveness in improving GUI testing.