The paper introduces WikiTrace, an evidence-linked memory system designed to mitigate memory-induced degradation in LLM agents by enabling verification of stored lessons against their original trajectories. External memory typically distills experience into concise insights provided without supporting evidence, causing agents to follow incorrect or misapplied lessons. WikiTrace addresses this by storing each insight linked to a summary of the trajectory's outcome and the raw trajectory itself. During inference on new tasks, agents retrieve all relevant insights and traverse these links using terminal tools to retrieve missing details or correct insights before acting. Evaluated across five agentic benchmarks, WikiTrace consistently outperforms existing memory approaches and a baseline using insights alone, demonstrating significant reduction in memory-induced degradation. The system's key innovation lies in its hierarchical evidence structure allowing agents to validate memory reliability dynamically during task execution.