Recommender systems must operate in dynamic environments where user preferences evolve rapidly across contexts, platforms, and temporal horizons. Conventional sequential recommenders struggle to adapt efficiently, often requiring extensive retraining or fine-tuning that incurs significant computational overhead. To overcome this issue, we suggest a meta-learned large language model recommender (Meta-LLMRec) that is aimed at quickly adapting to changing user behavior. Our model combines the concepts of meta-learning and token-level representations of language, allowing the recommender to learn transferable priors over heterogeneous tasks and dynamically learn more effectively with few gradient updates. In particular, we define the problem of preference prediction as a sequence-tosequence adaptation and optimize meta-initialized embeddings and Transformer encoders with a bilevel loss to reduce the generalization error on task distributions. Benchmark experiments of Meta-LLMRec against strong baselines, such as Amazon-Books and MovieLens-1M, show that it is more accurate and faster to adapt than baselines, by an average of 40% fewer gradient steps without compromising model stability. Ablation experiments also confirm that joint meta-learned priors, combined with tokenlevel semantics, perform significantly better in Cold-start scenarios, as well as in a change of temporal preference. These findings make Meta-LLMRec a scalable and effective paradigm in changing user preferences in realistic recommendation situations.