Fix bug in the case where the user produces a conversation by repeate…

…dly calling `agent.say`. Previously the agent would not observe the conversation in that case since the usual way of observing the conversation happens in the conversation scene, which this approach bypasses. The method that uses the conversation scene bypasses 'say', so this won't observe twice. This change just makes the behavior of 'say' closer to the behavior of the conversation scene.

PiperOrigin-RevId: 636322219
Change-Id: I90bc997e29b795e40604483caf9070a8df034393

concordia/agents/basic_agent.py

@@ -88,6 +88,7 @@ def __init__(
     self._user_controlled = user_controlled
     self._update_interval = update_interval
 
+    self._conversation_prefix = ''
     self._under_interrogation = False
     self._state_lock = threading.Lock()
 
@@ -287,7 +288,22 @@ def get_externality(externality):
 
     return output
 
+  def observe_latest(self, conversation: str):
+    # If the prefix is not found then `find` returns -1.
+    prefix_start_index = conversation.find(self._conversation_prefix)
+    if prefix_start_index >= 0:
+      # Get the part of the conversation the agent heard since their last turn.
+      start_index = prefix_start_index + len(self._conversation_prefix)
+      conversation_suffix = conversation[start_index:]
+      # Replace newline characters with commas.
+      conversation_suffix = conversation_suffix.replace('\n', ', ')
+      # Observe the new part of the conversation.
+      self.observe(conversation_suffix)
+      # Store the whole conversation thus far as the new prefix.
+      self._conversation_prefix = conversation
+
   def say(self, conversation: str) -> str:
+    self.observe_latest(conversation)
     convo_context = (
         f'{self._agent_name} is in the following'
         f' conversation:\n{conversation}\n'