diff --git a/pyproject.toml b/pyproject.toml
index 00b9f3a..26c8bda 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.9"
+version = "0.1.10"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -26,7 +26,7 @@ classifiers=[
 
 dependencies = [
     "accelerated-scan>=0.2.0",
-    "axial_positional_embedding>=0.3.5",
+    "axial_positional_embedding>=0.3.6",
     "einops>=0.8.0",
     "einx>=0.3.0",
     "hyper-connections>=0.1.8",
diff --git a/titans_pytorch/mac_transformer.py b/titans_pytorch/mac_transformer.py
index 5651108..51fef1e 100644
--- a/titans_pytorch/mac_transformer.py
+++ b/titans_pytorch/mac_transformer.py
@@ -593,11 +593,9 @@ def forward(
         # apply axial positional embedding
         # so intra and inter segment can be more easily discerned by the network
 
-        neural_mem_windows = ceil(seq_len_with_mem / neural_mem_segment_len)
+        pos_emb = self.axial_pos_emb.forward_with_seq_len(seq_len_with_mem, (neural_mem_segment_len,))
 
-        pos_emb = self.axial_pos_emb((neural_mem_windows, neural_mem_segment_len), flatten = True)
-
-        x = x + pos_emb[:seq_len_with_mem]
+        x = x + pos_emb
 
         # prep flex attention