diff --git a/skore-local-project/src/skore_local_project/project.py b/skore-local-project/src/skore_local_project/project.py
index cd2b707270..ef4657a116 100644
--- a/skore-local-project/src/skore_local_project/project.py
+++ b/skore-local-project/src/skore_local_project/project.py
@@ -170,7 +170,9 @@ def pickle(report: EstimatorReport | CrossValidationReport) -> tuple[str, bytes]
         The report is pickled without its cache, to avoid salting the hash.
         """
         reports = [report] + getattr(report, "estimator_reports_", [])
-        caches = [report_to_clear.__dict__.pop("_cache") for report_to_clear in reports]
+        caches = [report_to_clear._cache for report_to_clear in reports]
+
+        report.clear_cache()
 
         try:
             with io.BytesIO() as stream:
diff --git a/skore-local-project/tests/unit/test_project.py b/skore-local-project/tests/unit/test_project.py
index 918422172f..2e9d6f1b6f 100644
--- a/skore-local-project/tests/unit/test_project.py
+++ b/skore-local-project/tests/unit/test_project.py
@@ -126,36 +126,56 @@ def test_init(self, monkeypatch, tmp_path):
     def test_pickle_estimator_report(self, regression):
         # Pickle the report once, without any value in the cache
         assert not regression._cache
-        pickle_1 = Project.pickle(regression)
+        hash1, pickle1 = Project.pickle(regression)
         assert not regression._cache
 
         # Pickle the same report, but with values in the cache
         regression.cache_predictions()
 
         assert regression._cache
-        pickle_2 = Project.pickle(regression)
+        hash2, pickle2 = Project.pickle(regression)
         assert regression._cache
 
         # Make sure that the two pickles on the report are not affected by the cache
-        assert pickle_1 == pickle_2
+        assert (hash1, pickle1) == (hash2, pickle2)
+
+        # Make sure that pickles are not broken
+        with BytesIO(pickle1) as stream:
+            report1 = joblib.load(stream)
+
+        with BytesIO(pickle2) as stream:
+            report2 = joblib.load(stream)
+
+        report1.cache_predictions()
+        report2.cache_predictions()
 
     def test_pickle_cross_validation_report(self, cv_regression):
         reports = [cv_regression] + cv_regression.estimator_reports_
 
         # Pickle the report once, without any value in the cache
         assert not any(report._cache for report in reports)
-        pickle_1 = Project.pickle(cv_regression)
+        hash1, pickle1 = Project.pickle(cv_regression)
         assert not any(report._cache for report in reports)
 
         # Pickle the same report, but with values in the cache
         cv_regression.cache_predictions()
 
         assert any(report._cache for report in reports)
-        pickle_2 = Project.pickle(cv_regression)
+        hash2, pickle2 = Project.pickle(cv_regression)
         assert any(report._cache for report in reports)
 
         # Make sure that the two pickles on the report are not affected by the cache
-        assert pickle_1 == pickle_2
+        assert (hash1, pickle1) == (hash2, pickle2)
+
+        # Make sure that pickles are not broken
+        with BytesIO(pickle1) as stream:
+            report1 = joblib.load(stream)
+
+        with BytesIO(pickle2) as stream:
+            report2 = joblib.load(stream)
+
+        report1.cache_predictions()
+        report2.cache_predictions()
 
     def test_init_with_envar(self, monkeypatch, tmp_path):
         monkeypatch.setenv("SKORE_WORKSPACE", str(tmp_path))