One of the key challenges in automated chemical synthesis planning is to propose diverse and reliable reactions. A common approach is to generate reactions using reaction templates, which represent a reaction as a fixed graph transformation. This enables accurate and interpretable predictions but can suffer from limited diversity. On the other hand, template-free methods increase diversity but can be prone to making trivial mistakes. Inspired by the efficacy of reaction templates, we propose Molecule Edit Graph Attention Network (MEGAN), a template-free model that encodes reaction as a sequence of graph edits. Our model achieves state-of-the-art results on a standard retrosynthesis benchmark without any manual rule encoding.