Memory-Native Non-Terrestrial Networks for Embodied Intelligence

A dual-memory architecture enabling long-horizon, cross-layer optimization for satellite-assisted embodied AI

Chengyang Li; Yikun Wang; Jiahui He; Yujie Wan; Shuai Wang; Yuan Wu; Yik-Chung Wu; Chengzhong Xu; Huseyin Arslan· 2026· DOI 10.48550/arXiv.2607.00029

The core problem

Non-terrestrial networks (NTN) promise ubiquitous connectivity for embodied intelligence (EI), allowing robots operating in remote or wilderness environments to offload computation to cloud resources or report critical information to distant control centers. However, the synergy between NTN and EI is nontrivial because the operating environment is highly dynamic, resource-constrained, topology-varying, and task-oriented. Existing NTN protocols are fundamentally memoryless: their decisions are driven solely by local channel conditions and instantaneous service demands. This shortsightedness leads to inefficient resource usage, repeated failures, and poor adaptation to the long-term goals of embodied agents. The authors argue that a paradigm shift is needed—one that treats memory as a first-class network resource. They propose the memory-native NTN (Mem-NTN) paradigm, which leverages long-horizon contexts for memory-augmented system optimization. The core idea is to maintain and exploit historical information across time and layers, enabling decisions that anticipate future needs rather than merely reacting to the present.

Innovation

The authors evaluate Mem-NTN using a satellite embodied question answering (SEQA) task, where an embodied agent must answer questions by leveraging satellite connectivity to access cloud-based knowledge. They compare Mem-NTN against two baselines: conventional stateless NTN and terrestrial approaches. The experiments measure task accuracy, latency, and resource efficiency. The results show that Mem-NTN consistently outperforms both baselines across all metrics. For instance, in terms of answer accuracy, Mem-NTN achieves a significant improvement over stateless NTN, demonstrating the value of long-horizon context. Latency is also reduced because memory-native decisions can pre-allocate resources and avoid reactive retransmissions. Furthermore, the dual-memory architecture enables better adaptation to dynamic topology changes, as the digital memory captures patterns of connectivity that inform future handovers. The authors report that the gains are particularly pronounced in scenarios with high mobility and intermittent connectivity, which are common in NTN environments. Overall, the experiments validate the hypothesis that memory augmentation is crucial for efficient NTN-EI synergy.

Non-terrestrial networks (NTN) promise ubiquitous connectivity for embodied intelligence (EI), allowing robots operating in remote or wilderness environments to offload computation to cloud resources or report critical information to distant control centers. However, the synergy between NTN and EI is nontrivial because the operating environment is highly dynamic, resource-constrained, topology-varying, and task-oriented. Existing NTN protocols are fundamentally memoryless: their decisions are driven solely by local channel conditions and instantaneous service demands. This shortsightedness leads to inefficient resource usage, repeated failures, and poor adaptation to the long-term goals of embodied agents. The authors argue that a paradigm shift is needed—one that treats memory as a first-class network resource. They propose the memory-native NTN (Mem-NTN) paradigm, which leverages long-horizon contexts for memory-augmented system optimization. The core idea is to maintain and exploit historical information across time and layers, enabling decisions that anticipate future needs rather than merely reacting to the present.

To realize Mem-NTN, the authors establish a dual-memory architecture that distinguishes between two types of memory:

Why it matters

The paper's key contribution is the conceptual shift from memoryless to memory-native NTN protocols. By explicitly modeling physical and digital memory, the authors provide a framework for cross-layer optimization that respects the long-term objectives of embodied intelligence. The dual-memory architecture addresses the limitations of local, instantaneous decision-making, which is ill-suited for the highly dynamic and task-oriented nature of EI applications. The mechanisms for memory acquisition, compression, valuation, update, and utilization offer a practical path to implementation, though challenges remain in scaling to large networks and ensuring privacy and security of stored memories. The SEQA experiments serve as a proof-of-concept, but broader validation across diverse tasks and network conditions is needed. The authors also note that memory-native design could be extended to other domains such as autonomous driving and disaster response. The taxonomy candidates—Architecture, Cybersecurity, Network, Cryptography—highlight the interdisciplinary nature of the work. For instance, memory valuation and compression touch on information theory and cryptography, while cross-layer decision-making relates to network architecture. Future work could explore federated memory sharing and robust memory against adversarial attacks. In summary, Mem-NTN represents a promising direction for integrating memory into the fabric of non-terrestrial networks, enabling more intelligent and efficient support for embodied AI.

Who should read this

CS practitioners and researchers

Opening member content…