Do agents need visual memory? Not Mem0/Supermemory

Are you referring to per-chat context graphs? Where each interaction updates a structured state (constraints, entities, dependencies), instead of relying on tools like Mem0/Supermemory that only persist raw memory.