Long-horizon large language model (LLM) agents are fundamentally constrained by their limited context window. As interactions evolve over extended periods, an increasing volume of information accumulates, including detailed tool descriptions, retrieved memories, and raw environmental feedback. This accumulation frequently pushes out other crucial information necessary for effective agent operation, leading to a degradation in performance. To address this core limitation, the paper introduces GenericAgent, a novel LLM agent designed to be token-efficient and self-evolving, specifically targeting the challenges posed by contextual accumulation in long-horizon tasks.