Recursive Joint-Embedding Predictive Agents – Recursive (RJEPA-R) introduces a unified agent architecture that fuses recursive symbolic control with latent world modeling under explicit stability and resource constraints. RJEPA-R combines: Recursive Language Models (RLMs) for unbounded-context reasoning and programmatic memory access, and Joint-Embedding Predictive Architectures (JEPAs) for action-conditioned latent dynamics and predictive planning, into a closed-loop system governed by control-theoretic convergence, information-regulated computation, and spectral alignment between reasoning and environment dynamics. Unlike conventional LLM-based agents, RJEPA-R: avoids fixed-depth or heuristic halting by using Input-to-State Stability (ISS)–based convergence criteria, suppresses wasteful reasoning through information-gain–per–compute regulation, and rejects ineffective plans via a spectral alignment gate that matches the temporal/frequency characteristics of control actions to the dynamics they address. The architecture explicitly decouples memory access, reasoning, and world dynamics, reconnecting them through a principled decision gate that transforms open-loop generation into closed-loop, goal-conditioned decision making. RJEPA-R is designed to scale beyond fixed context windows, reduce hallucination and drift, and support long-horizon reasoning, multimodal understanding, and planning. The framework is modular, model-agnostic, and suitable for deployment across language agents, multimodal systems, and embodied AI. This release provides the full academic specification, algorithmic formulation, and evaluation outline for RJEPA-R, intended as a foundation for reproducible research in scalable, grounded, and controllable agent architectures.Optimized Learning Interface for Virtual Interaction and Assistance.