Existing multi-objective optimization algorithms struggle to balance convergence and diversity, especially with high-dimensional objectives and irregular Pareto fronts (PFs). To address these limitations, this paper proposes MOEA/D-AMKT, a decomposition-based multi-task evolutionary algorithm that synergizes convergence enhancement and diversity preservation through collaborative task optimization. The algorithm introduces a bidirectional knowledge transfer mechanism and dynamic task selection based on historical rewards, allowing computational resources to be adaptively focused on the most promising tasks. Task 1 applies an improved Penalty-based Boundary Intersection (PBI) method to enhance convergence, while Task 2 employs a reference-point-based strategy to maintain diversity. Collaborative evolution is achieved through a bidirectional knowledge transfer mechanism: Task 1 contributes convergence knowledge to Task 2, while Task 2 injects elite non-dominated solutions into Task 1 each generation. A global elite archive is maintained throughout the process, preserving and disseminating high-quality solutions. Experiments on IMOP, ZDT, and IDTLZ benchmarks demonstrate that MOEA/D-AMKT significantly outperforms six advanced algorithms in terms of inverted generational distance (IGD) and hypervolume (HV). Results highlight its robustness in handling complex PFs and provide efficient solutions for resource-limited scenarios.