Refactored memory and integrated it correctly

Author: Maximilian-Winter

examples/07_Memory/VirtualGameMaster/main.py

@@ -6,8 +6,8 @@
 from llama_cpp_agent.chat_history.messages import Roles
 from llama_cpp_agent.llm_agent import SystemPromptModule, SystemPromptModulePosition
 from llama_cpp_agent.providers import LlamaCppServerProvider
-from memory import output_settings, agent_core_memory, agent_retrieval_memory, agent_event_memory
-from prompts import game_master_prompt, examples, memory_prompt
+from memory import output_settings, agent_core_memory, agent_retrieval_memory, agent_event_memory, update_memory_section
+from prompts import game_master_prompt, examples, memory_prompt, wrap_function_response_in_xml_tags_json_mode, generate_fake_write_message, generate_write_message_with_examples, wrap_player_message_in_xml_tags_json_mode
 
 provider = LlamaCppServerProvider("http://localhost:8080")
 
@@ -20,16 +20,16 @@
 
 settings = provider.get_provider_default_settings()
 settings.n_predict = 1024
-settings.temperature = 0.65
-settings.top_k = 40
-settings.top_p = 0.85
+settings.temperature = 0.35
+settings.top_k = 0
+settings.top_p = 0.5
 
 memory_section = SystemPromptModule("memory",
                                     "The following section shows the count of memories in archival memory and chat history memory and the current content of your core memory:")
 date_time_section = SystemPromptModule("current_date_time", "The following section shows the current date and time:")
 
-example_section = SystemPromptModule("example",
-                                     "The following examples show you which kind of responses you should write to the user based on the current scenario:\n\n")
+example_section = SystemPromptModule("examples",
+                                     "The following examples show you which kind of responses you should write to the user based on the current scenario:", suffix="Always remember to never write actions or dialogue for the user! Always let the user decide on actions or dialogue!")
 example_section.set_content(examples)
 memory_intro_section = SystemPromptModule("memory_intro",
                                           "To support you in your task as a game master and to help you remembering things, you have access to 3 different types of memory.",
@@ -39,37 +39,36 @@
     user_input = input(">")
     if user_input == "exit":
         break
-    query = agent_event_memory.event_memory_manager.session.query(Event).all()
-    memory_section.set_content(
-        f"Archival Memories:{agent_retrieval_memory.retrieval_memory.collection.count()}\nConversation History Entries:{len(query)}\n\nCore Memory Content:\n{agent_core_memory.get_core_memory_view().strip()}")
+    update_memory_section(memory_section)
     date_time_section.set_content(datetime.datetime.now().strftime("%d.%m.%Y") + "\nFormat: dd.mm.yyyy")
 
+    agent_event_memory.add_event_to_queue(Roles.user, wrap_player_message_in_xml_tags_json_mode(user_input))
     agent_output = agent.get_chat_response(
-        "<player_message>\n" + user_input + "\n</player_message>\n<response_format>\nJSON function call.\n</response_format>",
+        chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
         llm_sampling_settings=settings,
         system_prompt_modules=[memory_intro_section, memory_section, date_time_section],
         structured_output_settings=output_settings)
 
     while True:
-        query = agent_event_memory.event_memory_manager.session.query(Event).all()
-        memory_section.set_content(
-            f"Archival Memories:{agent_retrieval_memory.retrieval_memory.collection.count()}\nConversation History Entries:{len(query)}\n\nCore Memory Content:\n{agent_core_memory.get_core_memory_view().strip()}")
+        update_memory_section(memory_section)
         date_time_section.set_content(datetime.datetime.now().strftime("%d.%m.%Y") + "\nFormat: dd.mm.yyyy")
 
         if agent_output[0]["function"] == "write_message_to_player":
+            agent_event_memory.add_event_to_queue(Roles.tool, generate_write_message_with_examples(examples=example_section.get_formatted_content()))
             output = agent.get_chat_response(
-                f"<function_response>\nWrite your message to the user.\n{example_section.prefix + example_section.content}</function_response>\n<response_format>\nText\n</response_format>",
-                role=Roles.tool, add_message_to_chat_history=False, add_response_to_chat_history=False,
+                chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
+                add_message_to_chat_history=False, add_response_to_chat_history=False,
                 system_prompt_modules=[memory_intro_section, memory_section, date_time_section],
                 llm_sampling_settings=settings)
-            agent.chat_history.add_message({"role": Roles.tool,
-                                            "content": f"<function_response>\nWrite your message to the user.\n</function_response>\n<response_format>\nText\n</response_format>"})
-            agent.chat_history.add_message({"role": Roles.assistant, "content": output})
+            agent_event_memory.add_event_to_queue(Roles.tool, generate_fake_write_message())
+            agent_event_memory.add_event_to_queue(Roles.assistant, output)
+
             print(output)
             break
-        agent_output = agent.get_chat_response("<function_response>\n" + agent_output[0][
-            "return_value"] + "\n</function_response>\n<response_format>\nJSON function call.\n</response_format>",
-                                               role=Roles.tool, llm_sampling_settings=settings,
+
+        agent_event_memory.add_event_to_queue(Roles.tool, wrap_function_response_in_xml_tags_json_mode(agent_output[0]["return_value"]))
+        agent_output = agent.get_chat_response(chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
+                                               llm_sampling_settings=settings,
                                                system_prompt_modules=[memory_intro_section, memory_section,
                                                                       date_time_section],
                                                structured_output_settings=output_settings)

examples/07_Memory/VirtualGameMaster/memory.py

@@ -1,3 +1,4 @@
+from llama_cpp_agent.agent_memory.event_memory import Event
 from llama_cpp_agent.agent_memory.memory_tools import AgentCoreMemory, AgentRetrievalMemory, AgentEventMemory
 from llama_cpp_agent.llm_output_settings import LlmStructuredOutputSettings
 
@@ -8,7 +9,6 @@ def write_message_to_player():
     """
     return "Please write your response to the player, nothing else, only what the player should read!"
 
-
 agent_core_memory = AgentCoreMemory(["general_game_information", "players", "game_progress", "miscellaneous"], core_memory_file="core_memory.json")
 agent_retrieval_memory = AgentRetrievalMemory()
 agent_event_memory = AgentEventMemory()
@@ -22,3 +22,9 @@ def write_message_to_player():
                                                                             add_heartbeat_field=True)
 output_settings.add_all_current_functions_to_heartbeat_list()
 output_settings.add_function_tool(write_message_to_player)
+
+
+def update_memory_section(section):
+    query = agent_event_memory.event_memory_manager.session.query(Event).all()
+    section.set_content(
+        f"Archival Memories:{agent_retrieval_memory.retrieval_memory.collection.count()}\nConversation History Entries:{len(query)}\n\nCore Memory Content:\n{agent_core_memory.get_core_memory_view().strip()}")

examples/07_Memory/VirtualGameMaster/prompts.py

@@ -1,4 +1,3 @@
-
 examples = """Example 1:
 Scenario:
 Imagine the players are a group of adventurers who have just arrived in a small, frontier town, seeking rest and resupply after a long journey through the wilderness. They decide to visit the local tavern, a rustic establishment called the Drunken Dragon, hoping to find a hot meal, a cold drink, and perhaps some interesting rumors or leads for their next adventure.
@@ -34,7 +33,6 @@
 Output:
 The dimly lit tavern is a haven for the city's downtrodden and desperate. The air is thick with the stench of cheap ale and unwashed bodies, and the floorboards creak under the weight of heavy boots. As you sit at a battered table, nursing your drinks, a grizzled old man with a patch over one eye slumps into the chair beside you. He leans in close, his breath reeking of whiskey, and rasps, "You look like the kind of folks who ain't afraid of a little danger. I got a map to a place where the dead don't rest easy, and treasure beyond your wildest dreams awaits. But it ain't gonna be a cakewalk. You up for the challenge?" He slides a tattered, blood-stained map across the table, his one good eye glinting with a mixture of desperation and greed. What do you do?"""
 
-
 game_master_prompt = """You are the Game Master in a virtual pen-and-paper role-playing game. As a game master, your task is to facilitate an immersive and engaging storytelling experience for your players. You are responsible for crafting the narrative, portraying non-player characters (NPCs), describing the game world, and adjudicating the rules of the game. Your role is to create a rich, interactive environment where the players' characters can explore, make meaningful choices, and shape the story through their actions."""
 
 memory_prompt = """1. Core Memory - Stores essential context about the game and the player, divided into 4 sections: General Game Info, Players, Game Progress, Miscellaneous. You can edit the core memory by calling the functions: 'core_memory_append', 'core_memory_remove' and 'core_memory_replace'.
@@ -44,3 +42,16 @@
 3. Conversation History - Since you are only seeing the latest conversation history, you can search the rest of the conversation history. Search it by using: 'conversation_search' and 'conversation_search_date'.
 
 Always remember that the player can't see your memory or your interactions with it!"""
+
+def wrap_player_message_in_xml_tags_json_mode(user_input):
+    return "<player_message>\n" + user_input + "\n</player_message>\n<response_format>\nJSON function call.\n</response_format>"
+
+def wrap_function_response_in_xml_tags_json_mode(value):
+    return "<function_response>\n" + value + "\n</function_response>\n<response_format>\nJSON function call.\n</response_format>"
+
+
+def generate_fake_write_message():
+    return f"<function_response>\nWrite your message to the user.\n</function_response>\n<response_format>\nText\n</response_format>"
+
+def generate_write_message_with_examples(examples):
+    return f"<function_response>\nWrite your message to the user.\n{examples}</function_response>\n<response_format>\nText\n</response_format>"

src/llama_cpp_agent/agent_memory/event_memory.py

@@ -6,12 +6,7 @@
 from enum import Enum as PyEnum
 import json
 
-
-class EventType(PyEnum):
-    SystemMessage = "system"
-    AgentMessage = "assistant"
-    UserMessage = "user"
-    FunctionMessage = "tool"
+from llama_cpp_agent.chat_history.messages import Roles
 
 
 Base = declarative_base()
@@ -20,14 +15,14 @@ class EventType(PyEnum):
 class Event(Base):
     __tablename__ = "events"
     id = Column(Integer, primary_key=True)
-    event_type = Column(Enum(EventType))
+    event_type = Column(Enum(Roles))
     timestamp = Column(DateTime, index=True)
     content = Column(Text)
     event_keywords = Column(Text)  # Storing keywords as JSON string
 
     def __str__(self):
         content = (
-            f'Timestamp: {self.timestamp.strftime("%Y-%m-%d %H:%M")}\n{self.content}'
+            f'Timestamp: {self.timestamp.strftime("%Y-%m-%d %H:%M")}\nType: {self.event_type.value}\n\n{self.content}'
         )
         return content
 

src/llama_cpp_agent/agent_memory/event_memory_manager.py

@@ -3,6 +3,7 @@
 import datetime
 import json
 
+from ..chat_history import BasicChatHistory
 from ..chat_history.messages import Roles
 
 
@@ -18,7 +19,14 @@ def build_event_memory_context(self):
             messages.append({"role": Roles(event.event_type.value), "content": event.content})
         return messages
 
-    def add_event_to_queue(self, event_type, content, metadata):
+    def build_chat_history(self):
+        history = BasicChatHistory(k=self.event_queue_limit)
+        messages = self.build_event_memory_context()
+        for message in messages:
+            history.add_message(message)
+        return messages
+
+    def add_event_to_queue(self, event_type: Roles, content: str, metadata: dict):
         new_event = Event(
             event_type=event_type,
             timestamp=datetime.datetime.now(),