diff --git a/runtime/mount/python_lib/lplots/h5/h5ad/process_message.py b/runtime/mount/python_lib/lplots/h5/h5ad/process_message.py
index 3b96253..38fedd8 100644
--- a/runtime/mount/python_lib/lplots/h5/h5ad/process_message.py
+++ b/runtime/mount/python_lib/lplots/h5/h5ad/process_message.py
@@ -1,4 +1,5 @@
 import asyncio
+import json
 from collections.abc import Awaitable, Callable
 from typing import Any
 
@@ -19,6 +20,8 @@
 
 alignment_is_running = False
 
+init_data_cache: dict[tuple, dict[str, Any]] = {}
+
 
 async def process_h5ad_request(
     msg: dict[str, Any],
@@ -84,34 +87,67 @@ async def process_h5ad_request(
         if init_obs_key is None and init_var_key is None and len(possible_obs_keys) > 0:
             init_obs_key = possible_obs_keys[0]
 
-        obsm = None
-        index = None
-        recomputed_index = False
-        filters = None
-        if init_obsm_key is not None:
-            filters = msg.get("filters")
-            obsm, index, recomputed_index = get_obsm(
-                obj_id, adata, init_obsm_key, filters, max_visualization_cells
-            )
+        filters = msg.get("filters")
 
-        obs = None
-        unique_obs = None
-        nrof_obs = None
-        counts = None
-        if init_obs_key is not None and init_obs_key in adata.obs:
-            obs, (unique_obs, counts), nrof_obs = get_obs(
-                obj_id, adata, init_obs_key, max_visualization_cells
-            )
+        cache_key = (
+            obj_id,
+            init_obsm_key,
+            json.dumps(filters, sort_keys=True),
+            init_obs_key,
+            init_var_key,
+            int(max_visualization_cells),
+        )
 
-        gene_column = None
-        if (
-            init_var_key is not None
-            and init_obs_key is None
-            and init_var_key in adata.var_names
-        ):
-            gene_column = get_obs_vector(obj_id, adata, init_var_key)
+        cached_value = init_data_cache.get(cache_key)
+
+        if cached_value is None:
+            obsm = None
+            index = None
+            recomputed_index = False
+            if init_obsm_key is not None:
+                obsm, index, recomputed_index = get_obsm(
+                    obj_id, adata, init_obsm_key, filters, max_visualization_cells
+                )
+
+            obs = None
+            unique_obs = None
+            nrof_obs = None
+            counts = None
+            if init_obs_key is not None and init_obs_key in adata.obs:
+                obs, (unique_obs, counts), nrof_obs = get_obs(
+                    obj_id, adata, init_obs_key, max_visualization_cells
+                )
+
+            gene_column = None
+            if (
+                init_var_key is not None
+                and init_obs_key is None
+                and init_var_key in adata.var_names
+            ):
+                gene_column = get_obs_vector(obj_id, adata, init_var_key)
+
+            var_index, var_names = get_var_index(obj_id, adata)
+
+            cached_value = {
+                "init_recomputed_index": recomputed_index,
+                "init_obsm_values": obsm.tolist() if obsm is not None else None,
+                "init_obsm_index": index.tolist() if index is not None else None,
+                "init_obsm_filters": filters,
+                "init_obs_values": obs.tolist() if obs is not None else None,
+                "init_obs_unique_values": (
+                    unique_obs.tolist() if unique_obs is not None else None
+                ),
+                "init_obs_counts": counts.tolist() if counts is not None else None,
+                "init_obs_nrof_values": nrof_obs,
+                "init_var_index": var_index.tolist(),
+                "init_var_names": (var_names.tolist() if var_names is not None else None),
+                "init_var_values": (
+                    gene_column.tolist() if gene_column is not None else None
+                ),
+                "init_var_key": init_var_key if init_var_key is not None else None,
+            }
 
-        var_index, var_names = get_var_index(obj_id, adata)
+            init_data_cache[cache_key] = cached_value
 
         global alignment_is_running
 
@@ -134,32 +170,15 @@ async def process_h5ad_request(
                     # init state with these
                     "init_obs_key": init_obs_key,
                     "init_obsm_key": init_obsm_key,
-                    "init_recomputed_index": recomputed_index,
-                    "init_obsm_values": obsm.tolist() if obsm is not None else None,
-                    "init_obsm_index": index.tolist() if index is not None else None,
-                    "init_obsm_filters": filters,
-                    "init_obs_values": obs.tolist() if obs is not None else None,
-                    "init_obs_unique_values": (
-                        unique_obs.tolist() if unique_obs is not None else None
-                    ),
-                    "init_obs_counts": counts.tolist() if counts is not None else None,
-                    "init_obs_nrof_values": nrof_obs,
-                    # var info
-                    "init_var_index": var_index.tolist(),
-                    "init_var_names": (
-                        var_names.tolist() if var_names is not None else None
-                    ),
-                    # var color by info
-                    "init_var_values": (
-                        gene_column.tolist() if gene_column is not None else None
-                    ),
-                    "init_var_key": init_var_key if init_var_key is not None else None,
+                    **cached_value,
                     # alignment info
                     "alignment_is_running": alignment_is_running,
                     # views info
                     "init_views": adata.uns.get("latch_views", []),
                     # images
                     "init_images": adata.uns.get("latch_images", {}),
+                    "cache_key": cache_key,
+                    "was_cached": cached_value is not None,
                 }
             },
         }