diff --git a/features/steps/repositories_coverage_r2_steps.py b/features/steps/repositories_coverage_r2_steps.py
index 149f8f014..296bd933f 100644
--- a/features/steps/repositories_coverage_r2_steps.py
+++ b/features/steps/repositories_coverage_r2_steps.py
@@ -18,7 +18,7 @@ from behave import given, then, when
 from behave.runner import Context
 from sqlalchemy import create_engine
 from sqlalchemy.exc import OperationalError
-from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.orm import Session, scoped_session, sessionmaker
 
 from cleveragents.core.exceptions import DatabaseError
 from cleveragents.domain.models.core.checkpoint import Checkpoint, CheckpointMetadata
@@ -234,7 +234,20 @@ def step_fresh_db(context: Context) -> None:
     engine = create_engine("sqlite:///:memory:", echo=False)
     Base.metadata.create_all(engine)
     context.r2_engine = engine
-    context.r2_session_factory = sessionmaker(bind=engine)
+    # Use scoped_session so that every ``factory()`` call within the
+    # same thread returns the *same* Session instance.  With plain
+    # ``sessionmaker``, each ``factory()`` call creates a new Session.
+    # SQLite in-memory uses ``SingletonThreadPool`` (one connection per
+    # thread), so all sessions share the same connection.  When a
+    # session created inside a repository method goes out of scope,
+    # Python's garbage collector may close it, issuing an implicit
+    # ROLLBACK on the shared connection — wiping flushed-but-uncommitted
+    # rows written by *other* sessions.  Under high memory pressure
+    # (e.g. 32 parallel worker processes) GC fires often enough to
+    # cause intermittent data loss between ``flush()`` and ``commit()``.
+    # ``scoped_session`` avoids the problem entirely: one Session lives
+    # for the whole scenario, so there is no premature close/rollback.
+    context.r2_session_factory = scoped_session(sessionmaker(bind=engine))
 
     # Pre-create repos used by multiple scenarios
     context.r2_skill_repo = SkillRepository(