Finding and fixing software bugs is a tedious, time-consuming process for developers. Imagine an AI assistant that could automatically identify and repair these bugs, saving countless hours and resources. While this sounds like science fiction, Large Language Models (LLMs) are making significant strides in automated program repair (APR). However, current LLM-based APR methods often hit a roadblock: memory inefficiency. They struggle to handle the vast search space of potential code fixes, leading to crashes and reduced effectiveness. Researchers have explored techniques like quantizing LLMs (reducing the precision of their internal calculations) and sequential beam search (processing potential fixes one at a time) to address the memory bottleneck. However, these methods have limitations and don't fully solve the problem. This is where a new approach called FLAMES comes in. FLAMES combines the power of LLMs with a clever search strategy called semantic-guided patch generation. Instead of blindly exploring all possible fixes, FLAMES uses feedback from test cases to guide the LLM toward more promising solutions. It starts by generating initial patches using a simple, memory-efficient method and then iteratively refines these patches based on how well they pass the tests. This targeted approach allows FLAMES to explore a larger number of potential fixes without overwhelming the system's memory. Experiments on real-world bugs from the Defects4J and HumanEval-Java datasets show that FLAMES not only significantly reduces memory consumption (by up to 83%) but also finds more correct fixes than existing techniques. It fixed more unique bugs, highlighting its ability to tackle complex issues that other methods miss. Importantly, FLAMES achieved these improvements while using fewer resources and less time than some of the current leading APR methods. While FLAMES represents a significant advance in automated program repair, there's still room for improvement. Future research could focus on refining the reward function used to guide the search process and explore ways to balance the exploration of new fixes with the exploitation of promising candidates. Nonetheless, FLAMES offers a glimpse into a future where AI could significantly streamline the software development process, allowing developers to focus on more creative and strategic tasks.