Future wireless networks will be artificial intelligence (AI) native and highly heterogeneous, where 5G/6G, next generation Wi-Fi, low-altitude drone networks, and satellite networks, have to coexist in the crowded gigahertz and tens of gigahertz frequency bands. This introduces grave challenges in radio resource sharing due to fundamental technical specification discrepancies and complex time-varying interference patterns. As a remedy, we propose a learn-to-share (L2S) framework—a distributed multi-agent paradigm for spectrum sharing among multiple radio access technologies (RATs) that will coexist in the 6G era. In this framework, user terminals (UTs) that are mutually interfering with each other, regardless of which RAT they use, can operate as intelligent agents capable of autonomously making decisions on spectrum access through local sensing and learning, without centralized coordination. We formulate this spectrum sharing problem as a decentralized partially observable Markov decision process (Dec-POMDP) and design a spectral-temporal attention and recurrence (STAR) algorithm as the cognitive engine. STAR decomposes individual agent decisions into three tightly coupled subtasks: channel selection via Transformer networks that capture spatial interference correlations, threshold adaptation through bidirectional long short-term memory (LSTM) networks that model temporal channel dynamics, and waiting time scheduling using predictive LSTM planners that intelligently avoid collisions. Through extensive OMNeT++-based simulations, we demonstrate that STAR significantly outperforms traditional approaches: achieving 73.5% improvement in access success rate, 30.6% enhancement in fairness (Jain’s index), and 88.1% increase in system throughput compared with the conventional listen-before-talk (LBT) mechanism, while requiring 50.6% fewer convergence episodes than the deep recurrent Q-network (DRQN) baseline. Moreover, compared with the mixing Q-network (QMIX), a representative centralized training with decentralized execution (CTDE) baseline, our algorithm reduces the computing power needed per UT by 21.9% while achieving comparable access success rate and system throughput as a fully decentralized multi-agent framework.