diff --git a/britney.conf b/britney.conf
index 5c62df2..434d757 100644
--- a/britney.conf
+++ b/britney.conf
@@ -105,3 +105,4 @@ ADT_SUCCESS_BOUNTY     = 3
 ADT_BASELINE           = reference
 ADT_RETRY_URL_MECH     = run_id
 ADT_RETRY_OLDER_THAN   = 1
+ADT_REFERENCE_MAX_AGE  = 7
diff --git a/britney2/policies/autopkgtest.py b/britney2/policies/autopkgtest.py
index 1e687c0..db8703a 100644
--- a/britney2/policies/autopkgtest.py
+++ b/britney2/policies/autopkgtest.py
@@ -159,6 +159,12 @@ class AutopkgtestPolicy(BasePolicy):
             # Make adt_baseline optional
             setattr(self.options, 'adt_baseline', None)
 
+        if not hasattr(self.options, 'adt_reference_max_age'):
+            self.options.adt_reference_max_age = float('inf')
+        else:
+            self.options.adt_reference_max_age = \
+              int(self.options.adt_reference_max_age) * SECPERDAY
+
         # read the cached results that we collected so far
         if os.path.exists(self.results_cache_file):
             with open(self.results_cache_file) as f:
@@ -168,9 +174,10 @@ class AutopkgtestPolicy(BasePolicy):
 
             # The cache can contain results against versions of packages that
             # are not in any suite anymore. Strip those out, as we don't want
-            # to use those results.
+            # to use those results. Additionally, old references may be
+            # filtered out.
             if self.options.adt_baseline == 'reference':
-                self.filter_results_for_old_versions()
+                self.filter_old_results()
 
         else:
             self.logger.info('%s does not exist, re-downloading all results from swift', self.results_cache_file)
@@ -266,8 +273,13 @@ class AutopkgtestPolicy(BasePolicy):
                 result.append(self._now)
         return test_results
 
-    def filter_results_for_old_versions(self):
-        '''Remove results for old versions from the cache'''
+    def filter_old_results(self):
+        '''Remove results for old versions and reference runs from the cache.
+
+        For now, only delete reference runs. If we delete regular
+        results after a while, packages with lots of triggered tests may
+        never have all the results at the same time.
+'''
 
         test_results = self.test_results
         test_results_new = deepcopy(test_results)
@@ -275,6 +287,9 @@ class AutopkgtestPolicy(BasePolicy):
         for (trigger, trigger_data) in test_results.items():
             for (src, results) in trigger_data.items():
                 for (arch, result) in results.items():
+                    if trigger == REF_TRIG and \
+                      result[3] < self._now - self.options.adt_reference_max_age:
+                        del test_results_new[trigger][src][arch]
                     if not self.test_version_in_any_suite(src, result[1]):
                         del test_results_new[trigger][src][arch]
                 if len(test_results_new[trigger][src]) == 0:
@@ -874,6 +889,11 @@ class AutopkgtestPolicy(BasePolicy):
                 "Ignoring result for source %s and trigger %s as the tested version %s isn't found in any suite",
                 src, trigger, ver)
             return
+        if trigger == REF_TRIG and \
+           seen < self._now - self.options.adt_reference_max_age:
+            self.logger.debug(
+                "Ignoring reference result for source %s as it's too old", src)
+            return
 
         result = self.test_results.setdefault(trigger, {}).setdefault(
             src, {}).setdefault(arch, [Result.FAIL, None, '', 0])
diff --git a/tests/test_policy.py b/tests/test_policy.py
index e6d4f80..1b6286a 100644
--- a/tests/test_policy.py
+++ b/tests/test_policy.py
@@ -583,6 +583,7 @@ class TestAutopkgtestPolicy(unittest.TestCase):
             AutopkgtestPolicy,
             adt_amqp=self.amqp,
             adt_retry_older_than=1,
+            adt_baseline='reference',
             pkg_universe=simple_universe,
             inst_tester=simple_inst_tester)
         autopkgtest_policy_info = apply_src_policy(policy, PolicyVerdict.PASS, src_name)
@@ -594,6 +595,26 @@ class TestAutopkgtestPolicy(unittest.TestCase):
         amqp = self.read_amqp()
         assert len(amqp) == 0
 
+    def test_fail_reference_too_old(self):
+        src_name = 'pkg'
+        policy = initialize_policy(
+            'autopkgtest/fail-to-fail',
+            AutopkgtestPolicy,
+            adt_amqp=self.amqp,
+            adt_retry_older_than=1,
+            adt_baseline='reference',
+            adt_reference_max_age=1,
+            pkg_universe=simple_universe,
+            inst_tester=simple_inst_tester)
+        autopkgtest_policy_info = apply_src_policy(policy, PolicyVerdict.REJECTED_TEMPORARILY, src_name)
+        assert autopkgtest_policy_info[src_name + '/2.0'][ARCH][0] == 'RUNNING-REFERENCE'
+        assert autopkgtest_policy_info[src_name + '/2.0'][ARCH][1] == \
+            'data/autopkgtest/testing/amd64/' + src_name[0] + '/' + src_name + '/2/log.gz'
+        assert autopkgtest_policy_info[src_name + '/2.0'][ARCH][2] == \
+            'packages/' + src_name[0] + '/' + src_name + '/testing/amd64'
+        amqp = self.read_amqp()
+        assert 'migration-reference/0' in amqp
+
 
 if __name__ == '__main__':
     unittest.main()