Feature/test summarization against memory challenge (Significant-Grav…

…itas#3567)

Co-authored-by: Toran Bruce Richards <[email protected]>

autogpt/agent/agent.py

@@ -56,6 +56,10 @@ def __init__(
         cfg = Config()
         self.ai_name = ai_name
         self.memory = memory
+        self.summary_memory = (
+            "I was created."  # Initial memory necessary to avoid hilucination
+        )
+        self.last_memory_index = 0
         self.full_message_history = full_message_history
         self.next_action_count = next_action_count
         self.command_registry = command_registry

autogpt/llm/chat.py

@@ -12,6 +12,10 @@
 from autogpt.memory_management.store_memory import (
     save_memory_trimmed_from_context_window,
 )
+from autogpt.memory_management.summary_memory import (
+    get_newly_trimmed_messages,
+    update_running_summary,
+)
 
 cfg = Config()
 
@@ -36,10 +40,10 @@ def generate_context(prompt, relevant_memory, full_message_history, model):
         create_chat_message(
             "system", f"The current time and date is {time.strftime('%c')}"
         ),
-        create_chat_message(
-            "system",
-            f"This reminds you of these events from your past:\n{relevant_memory}\n\n",
-        ),
+        # create_chat_message(
+        #     "system",
+        #     f"This reminds you of these events from your past:\n{relevant_memory}\n\n",
+        # ),
     ]
 
     # Add messages from the full message history until we reach the token limit
@@ -81,21 +85,21 @@ def chat_with_ai(
             """
             model = cfg.fast_llm_model  # TODO: Change model from hardcode to argument
             # Reserve 1000 tokens for the response
-
             logger.debug(f"Token limit: {token_limit}")
             send_token_limit = token_limit - 1000
-            if len(full_message_history) == 0:
-                relevant_memory = ""
-            else:
-                recent_history = full_message_history[-5:]
-                shuffle(recent_history)
-                relevant_memories = permanent_memory.get_relevant(
-                    str(recent_history), 5
-                )
-                if relevant_memories:
-                    shuffle(relevant_memories)
-                relevant_memory = str(relevant_memories)
 
+            # if len(full_message_history) == 0:
+            #     relevant_memory = ""
+            # else:
+            #     recent_history = full_message_history[-5:]
+            #     shuffle(recent_history)
+            #     relevant_memories = permanent_memory.get_relevant(
+            #         str(recent_history), 5
+            #     )
+            #     if relevant_memories:
+            #         shuffle(relevant_memories)
+            #     relevant_memory = str(relevant_memories)
+            relevant_memory = ""
             logger.debug(f"Memory Stats: {permanent_memory.get_stats()}")
 
             (
@@ -105,33 +109,36 @@ def chat_with_ai(
                 current_context,
             ) = generate_context(prompt, relevant_memory, full_message_history, model)
 
-            while current_tokens_used > 2500:
-                # remove memories until we are under 2500 tokens
-                relevant_memory = relevant_memory[:-1]
-                (
-                    next_message_to_add_index,
-                    current_tokens_used,
-                    insertion_index,
-                    current_context,
-                ) = generate_context(
-                    prompt, relevant_memory, full_message_history, model
-                )
+            # while current_tokens_used > 2500:
+            #     # remove memories until we are under 2500 tokens
+            #     relevant_memory = relevant_memory[:-1]
+            #     (
+            #         next_message_to_add_index,
+            #         current_tokens_used,
+            #         insertion_index,
+            #         current_context,
+            #     ) = generate_context(
+            #         prompt, relevant_memory, full_message_history, model
+            #     )
 
             current_tokens_used += count_message_tokens(
                 [create_chat_message("user", user_input)], model
             )  # Account for user input (appended later)
 
+            current_tokens_used += 500  # Account for memory (appended later) TODO: The final memory may be less than 500 tokens
+
+            # Add Messages until the token limit is reached or there are no more messages to add.
             while next_message_to_add_index >= 0:
                 # print (f"CURRENT TOKENS USED: {current_tokens_used}")
                 message_to_add = full_message_history[next_message_to_add_index]
 
                 tokens_to_add = count_message_tokens([message_to_add], model)
                 if current_tokens_used + tokens_to_add > send_token_limit:
-                    save_memory_trimmed_from_context_window(
-                        full_message_history,
-                        next_message_to_add_index,
-                        permanent_memory,
-                    )
+                    # save_memory_trimmed_from_context_window(
+                    #     full_message_history,
+                    #     next_message_to_add_index,
+                    #     permanent_memory,
+                    # )
                     break
 
                 # Add the most recent message to the start of the current context,
@@ -146,6 +153,22 @@ def chat_with_ai(
                 # Move to the next most recent message in the full message history
                 next_message_to_add_index -= 1
 
+            # Insert Memories
+            if len(full_message_history) > 0:
+                (
+                    newly_trimmed_messages,
+                    agent.last_memory_index,
+                ) = get_newly_trimmed_messages(
+                    full_message_history=full_message_history,
+                    current_context=current_context,
+                    last_memory_index=agent.last_memory_index,
+                )
+                agent.summary_memory = update_running_summary(
+                    current_memory=agent.summary_memory,
+                    new_events=newly_trimmed_messages,
+                )
+                current_context.insert(insertion_index, agent.summary_memory)
+
             api_manager = ApiManager()
             # inform the AI about its remaining budget (if it has one)
             if api_manager.get_total_budget() > 0.0:

autogpt/memory_management/summary_memory.py

@@ -0,0 +1,112 @@
+import json
+from typing import Dict, List, Tuple
+
+from autogpt.config import Config
+from autogpt.llm.llm_utils import create_chat_completion
+
+cfg = Config()
+
+
+def get_newly_trimmed_messages(
+    full_message_history: List[Dict[str, str]],
+    current_context: List[Dict[str, str]],
+    last_memory_index: int,
+) -> Tuple[List[Dict[str, str]], int]:
+    """
+    This function returns a list of dictionaries contained in full_message_history
+    with an index higher than prev_index that are absent from current_context.
+
+    Args:
+        full_message_history (list): A list of dictionaries representing the full message history.
+        current_context (list): A list of dictionaries representing the current context.
+        last_memory_index (int): An integer representing the previous index.
+
+    Returns:
+        list: A list of dictionaries that are in full_message_history with an index higher than last_memory_index and absent from current_context.
+        int: The new index value for use in the next loop.
+    """
+    # Select messages in full_message_history with an index higher than last_memory_index
+    new_messages = [
+        msg for i, msg in enumerate(full_message_history) if i > last_memory_index
+    ]
+
+    # Remove messages that are already present in current_context
+    new_messages_not_in_context = [
+        msg for msg in new_messages if msg not in current_context
+    ]
+
+    # Find the index of the last message processed
+    new_index = last_memory_index
+    if new_messages_not_in_context:
+        last_message = new_messages_not_in_context[-1]
+        new_index = full_message_history.index(last_message)
+
+    return new_messages_not_in_context, new_index
+
+
+def update_running_summary(current_memory: str, new_events: List[Dict]) -> str:
+    """
+    This function takes a list of dictionaries representing new events and combines them with the current summary,
+    focusing on key and potentially important information to remember. The updated summary is returned in a message
+    formatted in the 1st person past tense.
+
+    Args:
+        new_events (List[Dict]): A list of dictionaries containing the latest events to be added to the summary.
+
+    Returns:
+        str: A message containing the updated summary of actions, formatted in the 1st person past tense.
+
+    Example:
+        new_events = [{"event": "entered the kitchen."}, {"event": "found a scrawled note with the number 7"}]
+        update_running_summary(new_events)
+        # Returns: "This reminds you of these events from your past: \nI entered the kitchen and found a scrawled note saying 7."
+    """
+    # Replace "assistant" with "you". This produces much better first person past tense results.
+    for event in new_events:
+        if event["role"].lower() == "assistant":
+            event["role"] = "you"
+            # Remove "thoughts" dictionary from "content"
+            content_dict = json.loads(event["content"])
+            if "thoughts" in content_dict:
+                del content_dict["thoughts"]
+            event["content"] = json.dumps(content_dict)
+        elif event["role"].lower() == "system":
+            event["role"] = "your computer"
+        # Delete all user messages
+        elif event["role"] == "user":
+            new_events.remove(event)
+
+    # This can happen at any point during execturion, not just the beginning
+    if len(new_events) == 0:
+        new_events = "Nothing new happened."
+
+    prompt = f'''Your task is to create a concise running summary of actions and information results in the provided text, focusing on key and potentially important information to remember.
+
+You will receive the current summary and the your latest actions. Combine them, adding relevant key information from the latest development in 1st person past tense and keeping the summary concise.
+
+Summary So Far:
+"""
+{current_memory}
+"""
+
+Latest Development:
+"""
+{new_events}
+"""
+'''
+
+    messages = [
+        {
+            "role": "user",
+            "content": prompt,
+        }
+    ]
+
+    current_memory = create_chat_completion(messages, cfg.fast_llm_model)
+
+    message_to_return = {
+        "role": "system",
+        "content": f"This reminds you of these events from your past: \n{current_memory}",
+    }
+
+    return message_to_return