diff --git a/src/main.zig b/src/main.zig
index 8c91fc7..2895d65 100644
--- a/src/main.zig
+++ b/src/main.zig
@@ -1,4 +1,6 @@
 const std = @import("std");
+const GeneralPurposeAllocator = std.heap.GeneralPurposeAllocator(.{});
+
 const core = @import("mach-core");
 const gpu = core.gpu;
 
@@ -51,18 +53,31 @@ const PrimitiveData = struct {
 
 pub const App = @This();
 
+gpa: GeneralPurposeAllocator,
+allocator: std.mem.Allocator,
+
 app_timer: core.Timer,
 title_timer: core.Timer,
 
 pipeline: *gpu.RenderPipeline,
 scene_uniform_buffer: *gpu.Buffer,
 scene_uniform_bind_group: *gpu.BindGroup,
-object_data: [3]ObjectData,
-primitives: [2]PrimitiveData,
+object_data: []ObjectData,
+primitives: []PrimitiveData,
 
 pub fn init(app: *App) !void {
     try core.init(.{});
 
+    // Set up a "general purpose allocator" that will handle allocations for
+    // the lifetime of the application, for which a more specific allocation
+    // strategy is not necessary.
+    //
+    // Here, `gpa` is an instance of this allocator, which handles the
+    // allocation logic, while `allocator` is the interface through which
+    // functions such as `alloc` and `free` are called.
+    app.gpa = GeneralPurposeAllocator{};
+    app.allocator = app.gpa.allocator();
+
     app.app_timer = try core.Timer.start();
     app.title_timer = try core.Timer.start();
 
@@ -124,7 +139,17 @@ pub fn init(app: *App) !void {
         // would see the back side of the triangles, which are culled.
         const rotation = zm.rotationY(std.math.tau / 2.0);
 
-        for (object_desc, &app.object_data) |desc, *object| {
+        // Allocate a slice to store as many ObjectData as we want to create.
+        //
+        // Using a slice instead of an array means that we could change how
+        // many object we want to render at compile time, however it requires
+        // allocating, and later freeing, memory to store the slice.
+        app.object_data = try app.allocator.alloc(ObjectData, object_desc.len);
+
+        // Note that for loops in Zig are a little different than you might
+        // know from other languages. They only look over arrays, slices,
+        // tuples and ranges, potentially multiple at once.
+        for (object_desc, app.object_data) |desc, *object| {
             const translation = zm.translation(desc.pos[0], desc.pos[1], desc.pos[2]);
             const model_matrix = zm.mul(rotation, translation);
 
@@ -136,6 +161,8 @@ pub fn init(app: *App) !void {
                 },
             );
 
+            // The `*object` syntax gets us a pointer to each element in the
+            // `object_data` slice, allowing us to override it within the loop.
             object.* = .{
                 .uniform_buffer = result.buffer,
                 .uniform_bind_group = result.bind_group,
@@ -145,6 +172,7 @@ pub fn init(app: *App) !void {
     }
 
     // Set up the primitives we want to render.
+    app.primitives = try app.allocator.alloc(PrimitiveData, 2);
     // Triangle
     app.primitives[0] = createPrimitive(
         &.{
@@ -246,13 +274,16 @@ pub fn deinit(app: *App) void {
     // Using `defer` here, so we can specify them
     // in the order they were created in `init`.
     defer core.deinit();
+    defer _ = app.gpa.deinit(); // TODO: Check for memory leaks?
     defer app.pipeline.release();
     defer app.scene_uniform_buffer.release();
     defer app.scene_uniform_bind_group.release();
+    defer app.allocator.free(app.object_data);
     defer for (app.object_data) |o| {
         o.uniform_buffer.release();
         o.uniform_bind_group.release();
     };
+    defer app.allocator.free(app.primitives);
     defer for (app.primitives) |p| {
         p.vertex_buffer.release();
         p.index_buffer.release();