diff --git a/metagpt/roles/architect.py b/metagpt/roles/architect.py
index f18430c1c..465beff05 100644
--- a/metagpt/roles/architect.py
+++ b/metagpt/roles/architect.py
@@ -34,6 +34,7 @@ class Architect(Role):
 
     def __init__(self, **kwargs) -> None:
         super().__init__(**kwargs)
+        self.enable_memory = False
         # Initialize actions specific to the Architect role
         self.set_actions([PrepareDocuments(send_to=any_to_str(self), context=self.context), WriteDesign])
 
@@ -54,6 +55,3 @@ class Architect(Role):
             self.rc.todo = self.actions[idx]
             return bool(self.rc.todo)
         return False
-
-    async def _observe(self, ignore_memory=False) -> int:
-        return await super()._observe(ignore_memory=True)
diff --git a/metagpt/roles/product_manager.py b/metagpt/roles/product_manager.py
index 4fa8cb041..4beab5366 100644
--- a/metagpt/roles/product_manager.py
+++ b/metagpt/roles/product_manager.py
@@ -32,7 +32,7 @@ class ProductManager(Role):
 
     def __init__(self, **kwargs) -> None:
         super().__init__(**kwargs)
-
+        self.enable_memory = False
         self.set_actions([PrepareDocuments(send_to=any_to_str(self)), WritePRD])
         self._watch([UserRequirement, PrepareDocuments])
         self.rc.react_mode = RoleReactMode.BY_ORDER
@@ -47,6 +47,3 @@ class ProductManager(Role):
             self.config.git_reinit = False
             self.todo_action = any_to_name(WritePRD)
         return bool(self.rc.todo)
-
-    async def _observe(self, ignore_memory=False) -> int:
-        return await super()._observe(ignore_memory=True)
diff --git a/metagpt/roles/project_manager.py b/metagpt/roles/project_manager.py
index e7b453c48..70bd3bf8b 100644
--- a/metagpt/roles/project_manager.py
+++ b/metagpt/roles/project_manager.py
@@ -34,7 +34,7 @@ class ProjectManager(Role):
 
     def __init__(self, **kwargs) -> None:
         super().__init__(**kwargs)
-
+        self.enable_memory = False
         self.set_actions([PrepareDocuments(send_to=any_to_str(self), context=self.context), WriteTasks])
         self._watch([UserRequirement, PrepareDocuments, WriteDesign])
 
@@ -52,6 +52,3 @@ class ProjectManager(Role):
             self.rc.todo = self.actions[idx]
             return bool(self.rc.todo)
         return False
-
-    async def _observe(self, ignore_memory=False) -> int:
-        return await super()._observe(ignore_memory=True)
diff --git a/metagpt/roles/qa_engineer.py b/metagpt/roles/qa_engineer.py
index c1e50347b..06befc5f7 100644
--- a/metagpt/roles/qa_engineer.py
+++ b/metagpt/roles/qa_engineer.py
@@ -43,6 +43,7 @@ class QaEngineer(Role):
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
+        self.enable_memory = False
 
         # FIXME: a bit hack here, only init one action to circumvent _think() logic,
         #  will overwrite _think() in future updates
@@ -188,8 +189,3 @@ class QaEngineer(Role):
             sent_from=self.profile,
             send_to=MESSAGE_ROUTE_TO_NONE,
         )
-
-    async def _observe(self, ignore_memory=False) -> int:
-        # This role has events that trigger and execute themselves based on conditions, and cannot rely on the
-        # content of memory to activate.
-        return await super()._observe(ignore_memory=True)
diff --git a/metagpt/roles/role.py b/metagpt/roles/role.py
index e92c9958e..4dc0e34d9 100644
--- a/metagpt/roles/role.py
+++ b/metagpt/roles/role.py
@@ -34,7 +34,14 @@ from metagpt.context_mixin import ContextMixin
 from metagpt.logs import logger
 from metagpt.memory import Memory
 from metagpt.provider import HumanProvider
-from metagpt.schema import AIMessage, Message, MessageQueue, SerializationMixin
+from metagpt.schema import (
+    AIMessage,
+    Message,
+    MessageQueue,
+    SerializationMixin,
+    Task,
+    TaskResult,
+)
 from metagpt.strategy.planner import Planner
 from metagpt.utils.common import any_to_name, any_to_str, role_raise_decorator
 from metagpt.utils.project_repo import ProjectRepo
@@ -134,6 +141,9 @@ class Role(SerializationMixin, ContextMixin, BaseModel):
     constraints: str = ""
     desc: str = ""
     is_human: bool = False
+    enable_memory: bool = (
+        True  # Stateless, atomic roles, or roles that use external memory can disable this to save memory.
+    )
 
     role_id: str = ""
     states: list[str] = []
@@ -400,11 +410,12 @@ class Role(SerializationMixin, ContextMixin, BaseModel):
             msg = response
         else:
             msg = AIMessage(content=response, cause_by=self.rc.todo, sent_from=self)
-        self.rc.memory.add(msg)
+        if self.enable_memory:
+            self.rc.memory.add(msg)
 
         return msg
 
-    async def _observe(self, ignore_memory=False) -> int:
+    async def _observe(self) -> int:
         """Prepare new messages for processing from the message buffer and other sources."""
         # Read unprocessed messages from the msg buffer.
         news = []
@@ -413,7 +424,7 @@ class Role(SerializationMixin, ContextMixin, BaseModel):
         if not news:
             news = self.rc.msg_buffer.pop_all()
         # Store the read messages in your own memory to prevent duplicate processing.
-        old_messages = [] if ignore_memory else self.rc.memory.get()
+        old_messages = [] if not self.enable_memory else self.rc.memory.get()
         self.rc.memory.add_batch(news)
         # Filter out messages of interest.
         self.rc.news = [