Explicit Language Memory for Long-Horizon Planning in Vision-Language-Action Models