diff --git a/.jules/bolt.md b/.jules/bolt.md
index ff54159..bfb9ef8 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -9,3 +9,7 @@
 ## 2026-06-20 - Fast Spatial Neighborhood Checks
 **Learning:** In hot paths doing spatial scanning (like SparseAttentionGraph), calling `libm::sqrt` for distance thresholds is a massive bottleneck. Moreover, using a squared distance optimization `d^2 < r^2` must be robust to NaN values.
 **Action:** Always optimize spatial neighborhood checks using squared distances (`eps_sq`) inside an inline loop. Explicitly check for invalid thresholds before the loop (e.g. `!(epsilon > 0.0)`), and use the `!(sum < eps_sq)` pattern for early exits to handle NaNs safely while matching standard Euclidean semantics without the `sqrt` overhead.
+
+## 2026-06-24 - Zero-Copy Backpropagation via Option::take()
+**Learning:** In `aether-core`, `Tensor` structs contain heap-allocated metadata (like shape and strides). Operations like `last_z.as_ref().unwrap().clone()` in `DenseLayer::backward` trigger redundant heap allocations and slow down the training loop, even when the data itself is reference-counted.
+**Action:** When a cached tensor is only needed once during the backward pass (e.g., `last_z` and `last_input`), use `Option::take()` instead of `as_ref().clone()`. This transfers ownership out of the cache without cloning the metadata, saving memory allocations during backpropagation.
diff --git a/crates/aether-core/src/ml/neural.rs b/crates/aether-core/src/ml/neural.rs
index 70ec2dc..335319f 100644
--- a/crates/aether-core/src/ml/neural.rs
+++ b/crates/aether-core/src/ml/neural.rs
@@ -14,7 +14,6 @@
 // ═══════════════════════════════════════════════════════════════════════════════
 //
 
-
 #![allow(dead_code)]
 
 #[cfg(feature = "alloc")]
@@ -263,14 +262,12 @@ impl DenseLayer {
     pub fn backward(&mut self, grad_output: &Tensor, config: &OptimizerConfig) -> Tensor {
         let last_z = self
             .last_z
-            .as_ref()
-            .expect("Forward must be called before backward")
-            .clone();
+            .take()
+            .expect("Forward must be called before backward");
         let last_input = self
             .last_input
-            .as_ref()
-            .expect("Forward must be called before backward")
-            .clone();
+            .take()
+            .expect("Forward must be called before backward");
 
         let act_deriv = self.activation.derivative(&last_z);
         let delta = grad_output.mul(&act_deriv);
@@ -409,8 +406,15 @@ impl MLP {
 
     /// Forward pass through all layers
     pub fn forward(&mut self, input: &Tensor) -> Tensor {
-        let mut current = input.clone();
-        for layer in &mut self.layers {
+        if self.layers.is_empty() {
+            return input.clone();
+        }
+
+        let mut iter = self.layers.iter_mut();
+        let first_layer = iter.next().unwrap();
+        let mut current = first_layer.forward(input);
+
+        for layer in iter {
             current = layer.forward(&current);
         }
         current