The rapid advancement of generative agents has enabled the simulation of increasingly realistic human behaviors in artificial societies, yet challenges persist in scalability, resource efficiency, and emotional depth. This study introduces a novel framework, named CogniSociety, to enhance the fidelity and coherence of generative agents in complex social environments through three key innovations. First, a shared vector database facilitates collective knowledge sharing among agents, reducing redundant computations and improving coordination efficiency. Second, an affective modeling approach, leveraging the TweetNLP platform, integrates emotional states into agent decision-making, enabling nuanced human-like interactions. Third, the Solo Performance Prompting (SPP) architecture simulates policy-making institutions by dynamically orchestrating multi-agent collaboration, enhancing the realism of governance in sandbox environments. Experiments demonstrate that these contributions collectively address scalability limitations, emotional authenticity, and organizational complexity, advancing the potential of generative agents for applications in virtual worlds, social simulations, and policy analysis. The work also underscores ethical considerations for responsible AI deployment in human-AI ecosystems.