test: add thinking-mode compatibility tests for LTM retrieval (#272)

Co-authored-by: gitikavj <gitikavj@amazon.com>

Author: notgitika

tests/bedrock_agentcore/memory/integrations/strands/test_agentcore_memory_session_manager.py

@@ -1888,3 +1888,99 @@ def test_mixed_conversational_and_blob_messages(self, batching_session_manager,
         # Verify both paths were used
         assert mock_memory_client.create_event.call_count == 1  # Conversational
         assert mock_memory_client.gmdp_client.create_event.call_count == 1  # Blob
+
+
+class TestThinkingModeCompatibility:
+    """Test that retrieve_customer_context injects memory inline, not as an assistant message.
+
+    When thinking is enabled on Claude, assistant messages must start with a thinking block
+    and the conversation must end with a user message. Injecting LTM as a separate assistant
+    message violates both constraints.
+    """
+
+    def test_retrieve_customer_context_does_not_append_assistant_message(
+        self, agentcore_config_with_retrieval, mock_memory_client
+    ):
+        """Test retrieved memory is injected into the user message, not as a new assistant message."""
+        mock_memory_client.retrieve_memories.return_value = [
+            {"content": {"text": "User prefers dark mode"}},
+            {"content": {"text": "User likes sushi"}},
+        ]
+
+        with patch(
+            "bedrock_agentcore.memory.integrations.strands.session_manager.MemoryClient",
+            return_value=mock_memory_client,
+        ):
+            with patch("boto3.Session") as mock_boto_session:
+                mock_session = Mock()
+                mock_session.region_name = "us-west-2"
+                mock_session.client.return_value = Mock()
+                mock_boto_session.return_value = mock_session
+
+                with patch(
+                    "strands.session.repository_session_manager.RepositorySessionManager.__init__", return_value=None
+                ):
+                    manager = AgentCoreMemorySessionManager(agentcore_config_with_retrieval)
+
+                    mock_agent = Mock()
+                    mock_agent.messages = [{"role": "user", "content": [{"text": "What are my preferences?"}]}]
+
+                    event = MessageAddedEvent(
+                        agent=mock_agent, message={"role": "user", "content": [{"text": "What are my preferences?"}]}
+                    )
+                    manager.retrieve_customer_context(event)
+
+                    # No new messages should be added — memory is inlined in the user message
+                    assert len(mock_agent.messages) == 1
+                    assert mock_agent.messages[-1]["role"] == "user"
+
+                    # Memory prepended, original query remains last
+                    content = mock_agent.messages[0]["content"]
+                    assert len(content) == 2
+                    assert "<retrieved_memory>" in content[0]["text"]
+                    assert content[1]["text"] == "What are my preferences?"
+
+    def test_retrieve_customer_context_no_assistant_message_multi_turn(
+        self, agentcore_config_with_retrieval, mock_memory_client
+    ):
+        """Test memory injection keeps last message as user in a multi-turn conversation."""
+        mock_memory_client.retrieve_memories.return_value = [
+            {"content": {"text": "User likes sushi"}},
+        ]
+
+        with patch(
+            "bedrock_agentcore.memory.integrations.strands.session_manager.MemoryClient",
+            return_value=mock_memory_client,
+        ):
+            with patch("boto3.Session") as mock_boto_session:
+                mock_session = Mock()
+                mock_session.region_name = "us-west-2"
+                mock_session.client.return_value = Mock()
+                mock_boto_session.return_value = mock_session
+
+                with patch(
+                    "strands.session.repository_session_manager.RepositorySessionManager.__init__", return_value=None
+                ):
+                    manager = AgentCoreMemorySessionManager(agentcore_config_with_retrieval)
+
+                    mock_agent = Mock()
+                    mock_agent.messages = [
+                        {"role": "user", "content": [{"text": "I love sushi"}]},
+                        {"role": "assistant", "content": [{"text": "That's great!"}]},
+                        {"role": "user", "content": [{"text": "What do I like to eat?"}]},
+                    ]
+
+                    event = MessageAddedEvent(
+                        agent=mock_agent, message={"role": "user", "content": [{"text": "What do I like to eat?"}]}
+                    )
+                    manager.retrieve_customer_context(event)
+
+                    # No new messages added
+                    assert len(mock_agent.messages) == 3
+                    assert mock_agent.messages[-1]["role"] == "user"
+
+                    # Memory injected into last user message
+                    content = mock_agent.messages[-1]["content"]
+                    assert len(content) == 2
+                    assert "<retrieved_memory>" in content[0]["text"]
+                    assert content[1]["text"] == "What do I like to eat?"