diff --git a/fme/ace/aggregator/inference/test_zonal_mean.py b/fme/ace/aggregator/inference/test_zonal_mean.py
index 6fc159cbb..9cc7f9771 100644
--- a/fme/ace/aggregator/inference/test_zonal_mean.py
+++ b/fme/ace/aggregator/inference/test_zonal_mean.py
@@ -176,6 +176,38 @@ def test_zonal_mean_time_coarsening(n_time):
         )
 
 
+def test_zonal_mean_time_coarsening_25_steps_per_window():
+    """Regression: forward_steps_in_memory=25 with factor=2 causes 13 vs 12
+    size mismatch.
+
+    The bug occurs when there is no buffer and i_time_start is not aligned to the
+    coarsening factor. E.g. i_time_start=1, 25 steps: original code uses
+    time_slice length (1+25)//2 - 0 = 13 but _coarsen_tensor(25 steps) returns 12.
+    We must start at i_time_start=1 so the first batch has no buffer.
+    """
+    n_sample, ny, nx = 3, 10, 20
+    n_time = 100
+    window_steps = 25
+    # Factor 2: first batch at i_time_start=1 has no buffer -> 13 vs 12 mismatch
+    agg = ZonalMeanAggregator(
+        zonal_mean,
+        n_timesteps=n_time,
+        zonal_mean_max_size=50,  # ceil(100/50)=2
+    )
+    assert agg.time_coarsening_factor == 2
+    # Start at 1 so first batch triggers the bug (no buffer, misaligned i_time_start)
+    for i_time_start in range(1, n_time, window_steps):
+        steps = min(window_steps, n_time - i_time_start)
+        arr = torch.arange(ny, dtype=torch.float32, device=get_device())
+        arr = arr[None, None, :, None].expand(n_sample, steps, ny, nx)
+        agg.record_batch(
+            {"a": arr}, {"a": arr}, {"a": arr}, {"a": arr}, i_time_start=i_time_start
+        )
+    assert agg._target_data is not None
+    assert agg._gen_data is not None
+    assert agg._target_data["a"].shape[1] == n_time // agg.time_coarsening_factor
+
+
 @pytest.mark.parametrize("zonal_mean_max_size", [4, 2**14, 2**16])
 def test_zonal_mean_time_coarsening_override(zonal_mean_max_size):
     n_time = 2**16
diff --git a/fme/ace/aggregator/inference/zonal_mean.py b/fme/ace/aggregator/inference/zonal_mean.py
index 9a2058c01..8529ebcf7 100644
--- a/fme/ace/aggregator/inference/zonal_mean.py
+++ b/fme/ace/aggregator/inference/zonal_mean.py
@@ -173,21 +173,17 @@ def record_batch(
         # if we have a buffer that means we didnt record the last batch
         if self._buffer_gen:
             start_idx = self.last_step
+            buffer_size = (
+                i_time_start + window_steps
+            ) - self.last_step * self.time_coarsening_factor
         else:
             start_idx = i_time_start // self.time_coarsening_factor
-
-        time_slice = slice(
-            start_idx,
-            (i_time_start + window_steps) // self.time_coarsening_factor,
-        )
-
-        self.last_step = (i_time_start + window_steps) // self.time_coarsening_factor
-
-        buffer_size = (
-            i_time_start + window_steps
-        ) - self.last_step * self.time_coarsening_factor
+            buffer_size = (i_time_start + window_steps) - (
+                (i_time_start + window_steps) // self.time_coarsening_factor
+            ) * (self.time_coarsening_factor)
 
         buffer = {}
+        time_slice = None
         for name, tensor in target_data.items():
             if name in self._target_data:
                 if self._buffer_target:
@@ -198,11 +194,19 @@ def record_batch(
                         ],
                         dim=self._time_dim,
                     )
-                self._target_data[name][:, time_slice, :] += self._coarsen_tensor(
-                    self._zonal_mean(tensor)
-                )
-                if buffer_size > 0:
-                    buffer[name] = tensor[:, -buffer_size:, :]
+                coarsened = self._coarsen_tensor(self._zonal_mean(tensor))
+                if time_slice is None:
+                    # Use actual coarsened size so slice matches when i_time_start is
+                    # misaligned with coarsening factor
+                    n_coarsened = coarsened.shape[self._time_dim]
+                    time_slice = slice(start_idx, start_idx + n_coarsened)
+                    self.last_step = start_idx + n_coarsened
+                    new_buffer_size = (
+                        i_time_start + window_steps
+                    ) - self.last_step * self.time_coarsening_factor
+                self._target_data[name][:, time_slice, :] += coarsened
+                if new_buffer_size > 0:
+                    buffer[name] = tensor[:, -new_buffer_size:, :]
         self._buffer_target = buffer
 
         buffer = {}
@@ -216,11 +220,10 @@ def record_batch(
                         ],
                         dim=self._time_dim,
                     )
-                self._gen_data[name][:, time_slice, :] += self._coarsen_tensor(
-                    self._zonal_mean(tensor)
-                )
-                if buffer_size > 0:
-                    buffer[name] = tensor[:, -buffer_size:, :]
+                coarsened = self._coarsen_tensor(self._zonal_mean(tensor))
+                self._gen_data[name][:, time_slice, :] += coarsened
+                if new_buffer_size > 0:
+                    buffer[name] = tensor[:, -new_buffer_size:, :]
         self._buffer_gen = buffer
 
         self._n_batches[:, time_slice, :] += 1