diff --git a/python/activator/middleware_interface.py b/python/activator/middleware_interface.py
index 973198fc..899df2f1 100644
--- a/python/activator/middleware_interface.py
+++ b/python/activator/middleware_interface.py
@@ -70,12 +70,8 @@
 _log_trace3 = logging.getLogger("TRACE3.lsst." + __name__)
 _log_trace3.setLevel(logging.CRITICAL)  # Turn off by default.
 
-# The number of calib datasets to keep from previous runs, or one less than the
-# number of calibs that may be present *during* a run.
-base_keep_limit = int(os.environ.get("LOCAL_REPO_CACHE_SIZE", 3))-1
-# Multipliers to base_keep_limit for refcats and templates.
-refcat_factor = int(os.environ.get("REFCATS_PER_IMAGE", 4))
-template_factor = int(os.environ.get("PATCHES_PER_IMAGE", 4))
+# The number of calib datasets to keep, including the current run.
+base_keep_limit = int(os.environ.get("LOCAL_REPO_CACHE_SIZE", 3))
 # Whether or not to export to the central repo.
 do_export = bool(int(os.environ.get("DEBUG_EXPORT_OUTPUTS", '1')))
 # The number of arcseconds to pad the region in preloading spatial datasets.
@@ -168,25 +164,7 @@ def make_local_cache():
     cache : `activator.caching.DatasetCache`
         An empty cache with configured caching strategy and limits.
     """
-    return DatasetCache(
-        base_keep_limit,
-        cache_sizes={
-            # TODO: find an API that doesn't require explicit enumeration
-            "goodSeeingCoadd": template_factor * base_keep_limit,
-            "deepCoadd": template_factor * base_keep_limit,
-            "template_coadd": template_factor * base_keep_limit,
-            "uw_stars_20240524": refcat_factor * base_keep_limit,
-            "uw_stars_20240228": refcat_factor * base_keep_limit,
-            "uw_stars_20240130": refcat_factor * base_keep_limit,
-            "cal_ref_cat_2_2": refcat_factor * base_keep_limit,
-            "ps1_pv3_3pi_20170110": refcat_factor * base_keep_limit,
-            "gaia_dr3_20230707": refcat_factor * base_keep_limit,
-            "gaia_dr2_20200414": refcat_factor * base_keep_limit,
-            "atlas_refcat2_20220201": refcat_factor * base_keep_limit,
-            "the_monster_20240904": refcat_factor * base_keep_limit,
-            "the_monster_20250219": refcat_factor * base_keep_limit,
-        },
-    )
+    return DatasetCache(base_keep_limit)
 
 
 def _get_sasquatch_dispatcher():
@@ -406,15 +384,16 @@ def _define_dimensions(self):
                                                 "instrument": self.instrument.getName(),
                                                 })
 
-    def _mark_dataset_usage(self, refs: collections.abc.Iterable[lsst.daf.butler.DatasetRef]):
-        """Mark requested datasets in the cache.
+    def _cache_datasets(self, refs: collections.abc.Iterable[lsst.daf.butler.DatasetRef]):
+        """Add or mark requested datasets in the cache.
 
         Parameters
         ----------
         refs : iterable [`lsst.daf.butler.DatasetRef`]
-            The datasets to mark. Assumed to all fit inside the cache.
+            The datasets to cache. Assumed to all fit inside the cache.
         """
-        self.cache.update(refs)
+        evicted = self.cache.update(refs)
+        self.butler.pruneDatasets(evicted, disassociate=True, unstore=True, purge=True)
         try:
             self.cache.access(refs)
         except LookupError as e:
@@ -734,7 +713,7 @@ def _find_refcats(self, dataset_type, region):
                                      bind={"search_region": region},
                                      find_first=True,
                                      ),
-                      all_callback=self._mark_dataset_usage,
+                      # Don't cache refcats
                       ))
         if refcats:
             _log.debug("Found %d new refcat datasets from catalog '%s'.", len(refcats), dataset_type.name)
@@ -775,7 +754,7 @@ def _find_templates(self, dataset_type, region, physical_filter):
                            bind={"search_region": region},
                            find_first=True,
                            ),
-            all_callback=self._mark_dataset_usage,
+            # Don't cache templates
         ))
         if templates:
             _log.debug("Found %d new template datasets of type %s.", len(templates), dataset_type.name)
@@ -835,7 +814,7 @@ def query_calibs_by_date(butler, label):
         calibs = set(_filter_datasets(
             self.read_central_butler, self.butler,
             query_calibs_by_date,
-            all_callback=self._mark_dataset_usage,
+            all_callback=self._cache_datasets,
         ))
         if calibs:
             _log.debug("Found %d new calib datasets of type '%s'.", len(calibs), dataset_type.name)
@@ -872,7 +851,7 @@ def _find_generic_datasets(self, dataset_type, detector_id, physical_filter):
                            data_id=data_id,
                            find_first=True,
                            ),
-            all_callback=self._mark_dataset_usage,
+            all_callback=self._cache_datasets,
         ))
         if datasets:
             _log.debug("Found %d new datasets of type %s.", len(datasets), dataset_type.name)
@@ -912,7 +891,7 @@ def _find_init_outputs(self):
             run = runs.get_output_run(self.instrument, self._deployment, pipeline_file, self._day_obs)
             types = self._get_init_output_types(pipeline_file)
             # Output runs are always cleared after execution, so _filter_datasets would always warn.
-            # This also means the init-outputs don't need to be cached with _mark_dataset_usage.
+            # This also means the init-outputs don't need to be cached with _cache_datasets.
             query = _generic_query(types, collections=run)
             datasets.update(query(self.read_central_butler, "source datasets"))
         if not datasets: