diff --git a/priorConditionedAnnealing/noise.py b/priorConditionedAnnealing/noise.py
index 9247e05..e68f399 100644
--- a/priorConditionedAnnealing/noise.py
+++ b/priorConditionedAnnealing/noise.py
@@ -99,6 +99,7 @@ class Perlin_Noise():
         self.magic = 3.141592653589  # Axis offset, should be (kinda) irrational
         # We want to genrate samples, that approx ~N(0,1)
         self.normal_factor = 14/99
+        self.clear_cache_every = 1024
         self.reset()
 
     def __call__(self, shape=None):
@@ -107,6 +108,8 @@ class Perlin_Noise():
         self.index += 1
         noise = [self.noise([self.index*self.scale, self.magic*a]) / self.normal_factor
                  for a in range(shape[-1])]
+        if self.index % self.clear_cache_every == 0:
+            self.noise.cache = {}
         return th.Tensor(noise)
 
     def reset(self):
@@ -126,12 +129,16 @@ class Harmonic_Perlin_Noise():
                 octaves_arr += [1/(int_octaves+2)*(octaves-int_octaves)]
         octaves_arr = np.array(octaves_arr)
         self.octaves = octaves_arr / np.linalg.norm(octaves_arr)
+        self.clear_cache_every = 1024
         self.reset()
 
     def __call__(self, shape=None):
         if shape == None:
             shape = self.known_shape
         harmonics = [noise(shape)*self.octaves[i] for i, noise in enumerate(self.noises)]
+        if self.index % self.clear_cache_every == 0:
+            for i, noise in enumerate(self.noises):
+                noise.cache = {}
         return sum(harmonics)
 
     def reset(self):