diff --git a/pyproject.toml b/pyproject.toml
index e22d562..85e75d3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,6 +10,7 @@ requires-python = ">=3.11"
 dependencies = [
     "katpoint>=0.10.2",
     "katsdptelstate[rdb]>=1.0.0",
+    "obstore>=0.8.2",
     "tensorstore>=0.1.81",
     "xarray>=2025.9.1",
 ]
diff --git a/src/xarray_kat/datatree_factory.py b/src/xarray_kat/datatree_factory.py
index 491ee18..1ceef44 100644
--- a/src/xarray_kat/datatree_factory.py
+++ b/src/xarray_kat/datatree_factory.py
@@ -19,7 +19,7 @@
 )
 from xarray_kat.katdal_types import corrprod_to_autocorr
 from xarray_kat.multiton import Multiton
-from xarray_kat.stores.vis_weight_flag_store_factory import VisWeightFlagFactory
+from xarray_kat.tensorstores.vis_weight_flag_store_factory import VisWeightFlagFactory
 from xarray_kat.utils import corrprods_to_baseline_pols
 from xarray_kat.xkat_types import UvwSignConventionType, VanVleckLiteralType
 
diff --git a/src/xarray_kat/obstore/__init__.py b/src/xarray_kat/obstore/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/xarray_kat/obstore/adapted_array.py b/src/xarray_kat/obstore/adapted_array.py
new file mode 100644
index 0000000..4f0ef4d
--- /dev/null
+++ b/src/xarray_kat/obstore/adapted_array.py
@@ -0,0 +1,279 @@
+import multiprocessing
+from concurrent.futures import ThreadPoolExecutor
+from itertools import product
+from threading import Lock
+from typing import Tuple, TypedDict, cast
+
+import numpy as np
+import numpy.typing as npt
+from xarray.backends.common import BackendArray
+from xarray.core.indexing import (
+  ExplicitIndexer,
+  expanded_indexer,
+  integer_types,
+)
+
+from xarray_kat.xkat_types import ArchiveArrayMetadata
+
+# A selection over (time, frequency, corrprod)
+DimRangeType = Tuple[slice, slice, slice]
+
+
+class PreferredChunksType(TypedDict):
+  time: int
+  frequency: int
+  corrprod: int
+
+
+class VisFlagWeightData:
+  __slots__ = ("_vis", "_weight", "_flag")
+
+  _vis: npt.NDArray | None
+  _weight: npt.NDArray | None
+  _flag: npt.NDArray | None
+
+  def __init__(self):
+    self._vis = None
+    self._weight = None
+    self._flag = None
+
+  @property
+  def has_data(self) -> bool:
+    return self._vis is not None and self._weight is not None and self._flag is not None
+
+
+class VisFlagWeightGrid:
+  _vis_meta: ArchiveArrayMetadata
+  _weight_meta: ArchiveArrayMetadata
+  _channel_weight_meta: ArchiveArrayMetadata
+  _flag_meta: ArchiveArrayMetadata
+  _preferred_chunks: PreferredChunksType
+  _pool: ThreadPoolExecutor
+  _grid: npt.NDArray
+  _locks: npt.NDArray
+
+  def __init__(
+    self,
+    vis_meta: ArchiveArrayMetadata,
+    weight_meta: ArchiveArrayMetadata,
+    channel_weight_meta: ArchiveArrayMetadata,
+    flag_meta: ArchiveArrayMetadata,
+    preferred_chunks: PreferredChunksType,
+  ):
+    self._vis_meta = vis_meta
+    self._weight_meta = weight_meta
+    self._channel_weight_meta = channel_weight_meta
+    self._flag_meta = flag_meta
+    self._preferred_chunks = preferred_chunks
+    self._pool = ThreadPoolExecutor(max_workers=multiprocessing.cpu_count())
+
+    if not (vis_meta.shape == weight_meta.shape == flag_meta.shape) or not (
+      vis_meta.shape[:2] == channel_weight_meta.shape
+    ):
+      raise ValueError("Archive Array shapes don't match")
+
+    self.shape = vis_meta.shape
+    ntime, nfreq, ncorrprod = self.shape
+
+    array_chunks = [
+      vis_meta.chunks,
+      weight_meta.chunks,
+      channel_weight_meta.chunks + (-1,),
+      flag_meta.chunks,
+    ]
+
+    time_chunks, freq_chunks, cp_chunks = (max(c) for c in zip(*array_chunks))
+    time_chunks = max(preferred_chunks.get("time", time_chunks), time_chunks)
+    freq_chunks = max(preferred_chunks.get("frequency", freq_chunks), freq_chunks)
+    cp_chunks = max(preferred_chunks.get("corrprod", cp_chunks), cp_chunks)
+    self.chunks = (time_chunks, freq_chunks, cp_chunks)
+
+    ntime_chunks, rem = divmod(ntime, time_chunks)
+    ntime_chunks += int(rem != 0)
+    nfreq_chunks, rem = divmod(nfreq, freq_chunks)
+    nfreq_chunks += int(rem != 0)
+    ncorrprod_chunks, rem = divmod(ncorrprod, cp_chunks)
+    ncorrprod_chunks += int(rem != 0)
+
+    shape = (ntime_chunks, nfreq_chunks, ncorrprod_chunks)
+    nelements = ntime_chunks * nfreq_chunks * ncorrprod_chunks
+    self._grid = np.asarray([VisFlagWeightData()] * nelements).reshape(shape)  # noqa: F841
+    self._locks = np.asarray([Lock()] * nelements).reshape(shape)  # noqa: F841
+
+  def _chunk_indexer(self, key):
+    """Returns an indexer of the form (chunk_id, source_indexer, target_indexer)
+    for each dimension in the grid"""
+    ndim = len(self.shape)
+    indexer = (
+      expanded_indexer(key, ndim) if not isinstance(key, ExplicitIndexer) else key
+    )
+
+    new_indexer = []
+
+    for index, chunk, size in zip(indexer, self.chunks, self.shape):
+      if isinstance(index, integer_types):
+        if index < 0:
+          index += size
+
+        chunk_index, source_index = divmod(index, chunk)
+
+        new_indexer.append([(chunk_index, source_index, index)])
+      elif isinstance(index, slice):
+        if index.step not in {None, 1}:
+          raise NotImplementedError(
+            f"slice steps {index.step} other than 1 are not currently supported"
+          )
+
+        if (index_start := 0 if index.start is None else index.start) < 0:
+          index_start += size
+
+        if (index_stop := size if index.stop is None else index.stop) < 0:
+          index_stop += size
+
+        start_chunk, start_rem = divmod(index_start, chunk)
+        end_chunk, end_rem = divmod(index_stop, chunk)
+
+        # The index addresses a single chunk
+        if start_chunk == end_chunk:
+          new_indexer.append([(start_chunk, slice(0, index_stop - index_start), index)])
+        else:
+          # Multiple chunks case
+
+          # Add the start chunk
+          new_index = [
+            (
+              start_chunk,
+              slice(start_rem, chunk),
+              slice(index_start, (next_chunk := start_chunk + 1) * chunk),
+            )
+          ]
+
+          # Middle chunks
+          for c in range(next_chunk, end_chunk):
+            new_index.append((c, slice(0, chunk), slice(c * chunk, (c + 1) * chunk)))
+
+          if end_rem > 0:
+            # Final chunk
+            new_index.append(
+              (
+                end_chunk,
+                slice(0, end_rem),
+                slice(end_chunk * chunk, index_stop),
+              )
+            )
+
+          new_indexer.append(new_index)
+      elif isinstance(index, np.ndarray):
+        # Convert negative indices
+        index = np.where(index >= 0, index, index + size)
+        argsort = np.argsort(index)
+        sorted_index = index[argsort]
+        index_chunks = sorted_index // chunk
+        splits = np.where(np.ediff1d(index_chunks, to_begin=0) != 0)[0]
+        # Compute indices within each chunk
+        source_indices = np.split(sorted_index - (index_chunks * chunk), splits)
+        # Compute target indices for each chunk
+        target_indices = np.split(np.arange(argsort.size)[argsort], splits)
+        new_indexer.append(
+          (c[0].item(), si, ti)
+          for c, si, ti in zip(
+            np.split(index_chunks, splits), source_indices, target_indices
+          )
+        )
+      else:
+        raise TypeError(f"{type(index)} was not an integer, slice or numpy array")
+
+    return new_indexer
+
+  def __getitem__(self, key):
+    for index in product(*self._chunk_indexer(key)):
+      chunk, source_indices, target_indices = zip(*index)
+      self._maybe_retrieve_chunk(chunk, source_indices, target_indices)
+
+  def _maybe_retrieve_chunk(self, chunk, source_index, target_index):
+    def archive_array_paths(chunk_extents, meta: ArchiveArrayMetadata) -> list[str]:
+      paths = []
+      chunk_starts = (
+        tuple(range(s, e, c)) for (s, e), c in zip(chunk_extents, meta.chunks)
+      )
+      for chunk_start in product(*chunk_starts):
+        path_parts = "_".join(f"{c:05d}" for c in chunk_start)
+        path = f"{meta.prefix}/{meta.name}/{path_parts}.npy"
+        paths.append(path)
+      return paths
+
+    with self._locks[chunk]:
+      if not cast(VisFlagWeightData, self._grid[chunk]).has_data:
+        chunk_extents = tuple((c * s, (c + 1) * s) for c, s in zip(chunk, self.chunks))
+
+        paths = []
+        paths.extend(archive_array_paths(chunk_extents, self._flag_meta))
+        paths.extend(archive_array_paths(chunk_extents, self._vis_meta))
+        paths.extend(archive_array_paths(chunk_extents, self._weight_meta))
+        paths.extend(archive_array_paths(chunk_extents, self._channel_weight_meta))
+        print(paths)
+
+
+class VFWAdapter(BackendArray):
+  def __init__(self, array: str):
+    self.array = array
+
+
+if __name__ == "__main__":
+  from xarray_kat.utils import normalize_chunks
+
+  prefix = "12345-sdp-l0"
+  ntime = 100
+  nfreq = 32
+  ncorrprod = (7 * 7 // 2) * 4
+
+  chunk_info = {
+    "correlator_data": {
+      "prefix": prefix,
+      "dtype": "<c8",  # complex64 little-endian
+      "shape": (ntime, nfreq, ncorrprod),
+      "chunks": normalize_chunks((1, 8, ncorrprod), (ntime, nfreq, ncorrprod)),
+    },
+    "flags": {
+      "prefix": prefix,
+      "dtype": "|u1",  # uint8
+      "shape": (ntime, nfreq, ncorrprod),
+      "chunks": normalize_chunks((8, 8, ncorrprod), (ntime, nfreq, ncorrprod)),
+    },
+    "weights": {
+      "prefix": prefix,
+      "dtype": "|u1",  # uint8
+      "shape": (ntime, nfreq, ncorrprod),
+      "chunks": normalize_chunks((8, 8, ncorrprod), (ntime, nfreq, ncorrprod)),
+    },
+    "weights_channel": {
+      "prefix": prefix,
+      "dtype": "<f4",  # float32 little-endian
+      "shape": (ntime, nfreq),
+      "chunks": normalize_chunks((8, 8), (ntime, nfreq)),
+    },
+  }
+
+  dim_labels = ("time", "frequency", "corrprod")
+
+  meta = {
+    name: ArchiveArrayMetadata(
+      name,
+      0,
+      dim_labels[: len(value["shape"])],
+      cast(str, value["prefix"]),
+      cast(Tuple[Tuple[int, ...], ...], value["chunks"]),
+      cast(str, value["dtype"]),
+    )
+    for name, value in chunk_info.items()
+  }
+
+  grid = VisFlagWeightGrid(
+    meta["correlator_data"],
+    meta["weights"],
+    meta["weights_channel"],
+    meta["flags"],
+    {"time": 2, "frequency": 8, "corrprod": 4},
+  )
+
+  print(grid[slice(10, 28), np.array([6, 11, 7, 7, 10, 11, 8, 12])])
diff --git a/src/xarray_kat/obstore/lifecycle_manager.py b/src/xarray_kat/obstore/lifecycle_manager.py
new file mode 100644
index 0000000..96e4fd5
--- /dev/null
+++ b/src/xarray_kat/obstore/lifecycle_manager.py
@@ -0,0 +1,124 @@
+import weakref
+from itertools import pairwise, product
+from threading import Lock
+from typing import Any, Iterable
+
+import numpy as np
+import numpy.typing as npt
+from xarray.backends.common import BackendArray
+
+
+class LifeCycleManager:
+  def __init__(self):
+    # Keeps the object alive as long as it's in this set
+    self._keep_alive = {}
+    # Tracks which fields have been touched
+    self._touch_log = {}
+    # Protect access to the manager
+    self._lock = Lock()
+
+  def register(self, obj: Any, required_fields: Iterable[str]):
+    obj_id = id(obj)
+    with self._lock:
+      self._keep_alive[obj_id] = obj
+      self._touch_log[obj_id] = set(required_fields)
+      # print(f"Registered {obj_id}: Persistence active.")
+
+  def touch(self, obj: Any, field: str):
+    obj_id = id(obj)
+    with self._lock:
+      if obj_id in self._touch_log:
+        self._touch_log[obj_id].discard(field)
+
+        # If no fields are left, drop the strong reference
+        if not self._touch_log[obj_id]:
+          # print(f"All fields accessed for {obj_id}. Releasing to GC.")
+          del self._keep_alive[obj_id]
+          del self._touch_log[obj_id]
+
+
+# The Global Manager
+_MANAGER = LifeCycleManager()
+
+
+class VisFlagWeight:
+  __slots__ = ("_vis", "_flag", "_weight", "_lock", "__weakref__")
+
+  _lock: Lock
+  _vis: npt.NDArray | None
+  _flag: npt.NDArray | None
+  _weight: npt.NDArray | None
+
+  def __init__(self):
+    _MANAGER.register(self, ["vis", "flags", "weights"])
+    self._lock = Lock()
+    self._vis = None
+    self._weight = None
+    self._flag = None
+
+  def _retrieve(self):
+    with self._lock:
+      pass
+
+  @property
+  def vis(self) -> npt.NDArray:
+    _MANAGER.touch(self, "vis")
+    return self._vis
+
+  @property
+  def weight(self) -> npt.NDArray:
+    _MANAGER.touch(self, "weights")
+    return self._weight
+
+  @property
+  def flag(self) -> npt.NDArray:
+    _MANAGER.touch(self, "flags")
+    return self._flag
+
+
+vfw: weakref.WeakValueDictionary = weakref.WeakValueDictionary()
+vfw_lock: Lock = Lock()
+
+
+class VFWArray(BackendArray):
+  def __init__(self, shape, dtype, chunks):
+    self.shape = shape
+    self.dtype = dtype
+    from xarray_kat.utils import normalize_chunks
+
+    self.chunks = normalize_chunks(chunks, shape)
+    ranges = [np.concatenate(([0], np.cumsum(c))).tolist() for c in self.chunks]
+
+    for dim_coord_pairs in product(*(pairwise(r) for r in ranges)):
+      key = tuple(dim_coord_pairs)
+      with vfw_lock:
+        vfw.setdefault(key, VisFlagWeight())
+
+
+if __name__ == "__main__":
+  print(len(vfw))
+
+  def func():
+    A = VFWArray((100, 16, 4), np.float64, ((11,), (5,), (3,)))  # noqa: F841
+    B = VFWArray((100, 16, 4), np.bool, ((11,), (5,), (3,)))  # noqa: F841
+    print(len(vfw))
+    D = next(iter(vfw.values()))
+    D.vis
+    D.flag
+    D.weight
+
+  func()
+  print(len(vfw))
+  import gc
+
+  gc.collect()
+  print(len(vfw))
+
+  def clear_fn():
+    for v in vfw.values():
+      v.vis
+      v.flag
+      v.weight
+
+  clear_fn()
+  print(len(vfw))
diff --git a/src/xarray_kat/stores/base_store.py b/src/xarray_kat/tensorstores/base_store.py
similarity index 100%
rename from src/xarray_kat/stores/base_store.py
rename to src/xarray_kat/tensorstores/base_store.py
diff --git a/src/xarray_kat/stores/flag_store.py b/src/xarray_kat/tensorstores/flag_store.py
similarity index 100%
rename from src/xarray_kat/stores/flag_store.py
rename to src/xarray_kat/tensorstores/flag_store.py
diff --git a/src/xarray_kat/stores/http_store.py b/src/xarray_kat/tensorstores/http_store.py
similarity index 100%
rename from src/xarray_kat/stores/http_store.py
rename to src/xarray_kat/tensorstores/http_store.py
diff --git a/src/xarray_kat/stores/vfw_store.py b/src/xarray_kat/tensorstores/vfw_store.py
similarity index 100%
rename from src/xarray_kat/stores/vfw_store.py
rename to src/xarray_kat/tensorstores/vfw_store.py
diff --git a/src/xarray_kat/stores/vis_weight_flag_store_factory.py b/src/xarray_kat/tensorstores/vis_weight_flag_store_factory.py
similarity index 100%
rename from src/xarray_kat/stores/vis_weight_flag_store_factory.py
rename to src/xarray_kat/tensorstores/vis_weight_flag_store_factory.py
diff --git a/src/xarray_kat/stores/visibility_stores.py b/src/xarray_kat/tensorstores/visibility_stores.py
similarity index 100%
rename from src/xarray_kat/stores/visibility_stores.py
rename to src/xarray_kat/tensorstores/visibility_stores.py
diff --git a/src/xarray_kat/stores/weight_store.py b/src/xarray_kat/tensorstores/weight_store.py
similarity index 100%
rename from src/xarray_kat/stores/weight_store.py
rename to src/xarray_kat/tensorstores/weight_store.py