diff --git a/src/adapters/sqlite_repository.py b/src/adapters/sqlite_repository.py
index f2110b4..d713f6f 100644
--- a/src/adapters/sqlite_repository.py
+++ b/src/adapters/sqlite_repository.py
@@ -1504,13 +1504,23 @@ def save_llm_response(self, prompt_hash: str, response_json: str) -> None:
                 """
                 UPDATE llm_calls
                 SET response_json = ?
-                WHERE prompt_hash = ?
-                ORDER BY ts DESC
-                LIMIT 1
+                WHERE id = (
+                    SELECT id
+                    FROM llm_calls
+                    WHERE prompt_hash = ?
+                    ORDER BY ts DESC, id DESC
+                    LIMIT 1
+                )
                 """,
                 (response_json, prompt_hash),
             )
 
+            if cursor.rowcount == 0:
+                logger.warning(
+                    "llm_response_update_missing_call",
+                    prompt_hash=prompt_hash[:12],
+                )
+
             conn.commit()
             conn.close()
 
diff --git a/tests/test_sqlite_llm_cache.py b/tests/test_sqlite_llm_cache.py
new file mode 100644
index 0000000..6498b7a
--- /dev/null
+++ b/tests/test_sqlite_llm_cache.py
@@ -0,0 +1,62 @@
+"""Tests for SQLite LLM cache persistence."""
+
+from __future__ import annotations
+
+import sqlite3
+from datetime import UTC, datetime, timedelta
+
+from src.adapters.sqlite_repository import SQLiteRepository
+from src.domain.models import LLMCallMetadata
+
+
+def test_save_llm_response_updates_latest_call(tmp_path) -> None:
+    """Ensure the cached response is stored on the latest call row."""
+
+    db_path = tmp_path / "llm_cache.db"
+    repository = SQLiteRepository(str(db_path))
+
+    prompt_hash = "prompt-hash"
+    older_metadata = LLMCallMetadata(
+        message_id="message-1",
+        prompt_hash=prompt_hash,
+        model="gpt-5-nano",
+        tokens_in=10,
+        tokens_out=15,
+        cost_usd=0.001,
+        latency_ms=1200,
+        cached=False,
+        ts=datetime.now(tz=UTC) - timedelta(minutes=5),
+    )
+    newer_metadata = LLMCallMetadata(
+        message_id="message-2",
+        prompt_hash=prompt_hash,
+        model="gpt-5-nano",
+        tokens_in=12,
+        tokens_out=18,
+        cost_usd=0.0015,
+        latency_ms=800,
+        cached=False,
+        ts=datetime.now(tz=UTC),
+    )
+
+    repository.save_llm_call(older_metadata)
+    repository.save_llm_call(newer_metadata)
+
+    cached_payload = '{"events": []}'
+    repository.save_llm_response(prompt_hash, cached_payload)
+
+    with sqlite3.connect(db_path) as conn:
+        conn.row_factory = sqlite3.Row
+        rows = conn.execute(
+            """
+            SELECT id, response_json, ts
+            FROM llm_calls
+            WHERE prompt_hash = ?
+            ORDER BY ts ASC
+            """,
+            (prompt_hash,),
+        ).fetchall()
+
+    assert len(rows) == 2
+    assert rows[0]["response_json"] is None
+    assert rows[1]["response_json"] == cached_payload