The proliferation of AI and deepfake technologies has enabled sophisticated fraud attacks, with financial institutions reporting over $6.8 billion in fraud losses in 2023. Traditional authentication methods often sacrifice user experience for security, leading to 73 % transaction abandonment rates. This paper proposes a novel adaptive dual-agent authentication framework that leverages Long-Term Memory (LTM) and Short-Term Memory (STM) AI agents for intelligent risk-based authentication decisions. Our framework integrates behavioral biometrics, machine learning-based risk scoring, and context-aware authentication orchestration to dynamically adjust security requirements. Through comprehensive evaluation on 10,620 transactions with 1,000 customer profiles, our framework achieves a 50.95 % fraud detection rate, successfully identifying over half of fraudulent attempts at the initial risk-assessment stage, with only a 9.43% false positive rate, outperforming traditional MFA by $\mathbf{1 5. 2 \%}$ in detection accuracy. Crucially, it processes decisions in submillisecond time (0.000012 s) and reduces authentication friction by 47.36 %, successfully balancing security and user experience. The framework's dual-agent architecture enables real-time behavioral analysis and adaptive authentication orchestration, making it suitable for high-volume digital banking environments where both security and user satisfaction are paramount.