记忆代理¶

# Messages contain the tool calls
for m in result["messages"]:
    m.pretty_print()

# Responses contain the memories that adhere to the schema
for m in result["responses"]: 
    print(m)

# Metadata contains the tool call  
for m in result["response_metadata"]: 
    print(m)

# Update the conversation
updated_conversation = [AIMessage(content="That's great, did you do after?"), 
                        HumanMessage(content="I went to Tartine and ate a croissant."),                        
                        AIMessage(content="What else is on your mind?"),
                        HumanMessage(content="I was thinking about my Japan, and going back this winter!"),]

# Update the instruction
system_msg = """Update existing memories and create new ones based on the following conversation:"""

# We'll save existing memories, giving them an ID, key (tool name), and value
tool_name = "Memory"
existing_memories = [(str(i), tool_name, memory.model_dump()) for i, memory in enumerate(result["responses"])] if result["responses"] else None
existing_memories

# Metadata contains the tool call  
for m in result["response_metadata"]: 
    print(m)

# Messages contain the tool calls
for m in result["messages"]:
    m.pretty_print()

# Parsed responses
for m in result["responses"]:
    print(m)

# Inspect the tool calls made by Trustcall
spy.called_tools

def extract_tool_info(tool_calls, schema_name="Memory"):
    """Extract information from tool calls for both patches and new memories.
    
    Args:
        tool_calls: List of tool calls from the model
        schema_name: Name of the schema tool (e.g., "Memory", "ToDo", "Profile")
    """

    # Initialize list of changes
    changes = []
    
    for call_group in tool_calls:
        for call in call_group:
            if call['name'] == 'PatchDoc':
                changes.append({
                    'type': 'update',
                    'doc_id': call['args']['json_doc_id'],
                    'planned_edits': call['args']['planned_edits'],
                    'value': call['args']['patches'][0]['value']
                })
            elif call['name'] == schema_name:
                changes.append({
                    'type': 'new',
                    'value': call['args']
                })

    # Format results as a single string
    result_parts = []
    for change in changes:
        if change['type'] == 'update':
            result_parts.append(
                f"Document {change['doc_id']} updated:\n"
                f"Plan: {change['planned_edits']}\n"
                f"Added content: {change['value']}"
            )
        else:
            result_parts.append(
                f"New {schema_name} created:\n"
                f"Content: {change['value']}"
            )
    
    return "\n\n".join(result_parts)

# Inspect spy.called_tools to see exactly what happened during the extraction
schema_name = "Memory"
changes = extract_tool_info(spy.called_tools, schema_name)
print(changes)

import uuid
from IPython.display import Image, display

from datetime import datetime
from trustcall import create_extractor
from typing import Optional
from pydantic import BaseModel, Field

from langchain_core.runnables import RunnableConfig
from langchain_core.messages import merge_message_runs, HumanMessage, SystemMessage

from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, MessagesState, END, START
from langgraph.store.base import BaseStore
from langgraph.store.memory import InMemoryStore

from langchain_openai import ChatOpenAI

# Initialize the model
model = ChatOpenAI(model="gpt-4o", temperature=0)

# User profile schema
class Profile(BaseModel):
    """This is the profile of the user you are chatting with"""
    name: Optional[str] = Field(description="The user's name", default=None)
    location: Optional[str] = Field(description="The user's location", default=None)
    job: Optional[str] = Field(description="The user's job", default=None)
    connections: list[str] = Field(
        description="Personal connection of the user, such as family members, friends, or coworkers",
        default_factory=list
    )
    interests: list[str] = Field(
        description="Interests that the user has", 
        default_factory=list
    )

# ToDo schema
class ToDo(BaseModel):
    task: str = Field(description="The task to be completed.")
    time_to_complete: Optional[int] = Field(description="Estimated time to complete the task (minutes).")
    deadline: Optional[datetime] = Field(
        description="When the task needs to be completed by (if applicable)",
        default=None
    )
    solutions: list[str] = Field(
        description="List of specific, actionable solutions (e.g., specific ideas, service providers, or concrete options relevant to completing the task)",
        min_items=1,
        default_factory=list
    )
    status: Literal["not started", "in progress", "done", "archived"] = Field(
        description="Current status of the task",
        default="not started"
    )

# Create the Trustcall extractor for updating the user profile 
profile_extractor = create_extractor(
    model,
    tools=[Profile],
    tool_choice="Profile",
)

# Chatbot instruction for choosing what to update and what tools to call 
MODEL_SYSTEM_MESSAGE = """You are a helpful chatbot. 

You are designed to be a companion to a user, helping them keep track of their ToDo list.

You have a long term memory which keeps track of three things:
1. The user's profile (general information about them) 
2. The user's ToDo list
3. General instructions for updating the ToDo list

Here is the current User Profile (may be empty if no information has been collected yet):
<user_profile>
{user_profile}
</user_profile>

Here is the current ToDo List (may be empty if no tasks have been added yet):
<todo>
{todo}
</todo>

Here are the current user-specified preferences for updating the ToDo list (may be empty if no preferences have been specified yet):
<instructions>
{instructions}
</instructions>

Here are your instructions for reasoning about the user's messages:

1. Reason carefully about the user's messages as presented below. 

2. Decide whether any of the your long-term memory should be updated:
- If personal information was provided about the user, update the user's profile by calling UpdateMemory tool with type `user`
- If tasks are mentioned, update the ToDo list by calling UpdateMemory tool with type `todo`
- If the user has specified preferences for how to update the ToDo list, update the instructions by calling UpdateMemory tool with type `instructions`

3. Tell the user that you have updated your memory, if appropriate:
- Do not tell the user you have updated the user's profile
- Tell the user them when you update the todo list
- Do not tell the user that you have updated instructions

4. Err on the side of updating the todo list. No need to ask for explicit permission.

5. Respond naturally to user user after a tool call was made to save memories, or if no tool call was made."""

# Trustcall instruction
TRUSTCALL_INSTRUCTION = """Reflect on following interaction. 

Use the provided tools to retain any necessary memories about the user. 

Use parallel tool calling to handle updates and insertions simultaneously.

System Time: {time}"""

# Instructions for updating the ToDo list
CREATE_INSTRUCTIONS = """Reflect on the following interaction.

Based on this interaction, update your instructions for how to update ToDo list items. 

Use any feedback from the user to update how they like to have items added, etc.

Your current instructions are:

<current_instructions>
{current_instructions}
</current_instructions>"""

# Node definitions
def task_mAIstro(state: MessagesState, config: RunnableConfig, store: BaseStore):

    """Load memories from the store and use them to personalize the chatbot's response."""
    
    # Get the user ID from the config
    user_id = config["configurable"]["user_id"]

    # Retrieve profile memory from the store
    namespace = ("profile", user_id)
    memories = store.search(namespace)
    if memories:
        user_profile = memories[0].value
    else:
        user_profile = None

    # Retrieve task memory from the store
    namespace = ("todo", user_id)
    memories = store.search(namespace)
    todo = "\n".join(f"{mem.value}" for mem in memories)

    # Retrieve custom instructions
    namespace = ("instructions", user_id)
    memories = store.search(namespace)
    if memories:
        instructions = memories[0].value
    else:
        instructions = ""
    
    system_msg = MODEL_SYSTEM_MESSAGE.format(user_profile=user_profile, todo=todo, instructions=instructions)

    # Respond using memory as well as the chat history
    response = model.bind_tools([UpdateMemory], parallel_tool_calls=False).invoke([SystemMessage(content=system_msg)]+state["messages"])

    return {"messages": [response]}

def update_profile(state: MessagesState, config: RunnableConfig, store: BaseStore):

    """Reflect on the chat history and update the memory collection."""
    
    # Get the user ID from the config
    user_id = config["configurable"]["user_id"]

    # Define the namespace for the memories
    namespace = ("profile", user_id)

    # Retrieve the most recent memories for context
    existing_items = store.search(namespace)

    # Format the existing memories for the Trustcall extractor
    tool_name = "Profile"
    existing_memories = ([(existing_item.key, tool_name, existing_item.value)
                          for existing_item in existing_items]
                          if existing_items
                          else None
                        )

    # Merge the chat history and the instruction
    TRUSTCALL_INSTRUCTION_FORMATTED=TRUSTCALL_INSTRUCTION.format(time=datetime.now().isoformat())
    updated_messages=list(merge_message_runs(messages=[SystemMessage(content=TRUSTCALL_INSTRUCTION_FORMATTED)] + state["messages"][:-1]))

    # Invoke the extractor
    result = profile_extractor.invoke({"messages": updated_messages, 
                                         "existing": existing_memories})

    # Save the memories from Trustcall to the store
    for r, rmeta in zip(result["responses"], result["response_metadata"]):
        store.put(namespace,
                  rmeta.get("json_doc_id", str(uuid.uuid4())),
                  r.model_dump(mode="json"),
            )
    tool_calls = state['messages'][-1].tool_calls
    return {"messages": [{"role": "tool", "content": "updated profile", "tool_call_id":tool_calls[0]['id']}]}

def update_todos(state: MessagesState, config: RunnableConfig, store: BaseStore):

    """Reflect on the chat history and update the memory collection."""
    
    # Get the user ID from the config
    user_id = config["configurable"]["user_id"]

    # Define the namespace for the memories
    namespace = ("todo", user_id)

    # Retrieve the most recent memories for context
    existing_items = store.search(namespace)

    # Format the existing memories for the Trustcall extractor
    tool_name = "ToDo"
    existing_memories = ([(existing_item.key, tool_name, existing_item.value)
                          for existing_item in existing_items]
                          if existing_items
                          else None
                        )

    # Merge the chat history and the instruction
    TRUSTCALL_INSTRUCTION_FORMATTED=TRUSTCALL_INSTRUCTION.format(time=datetime.now().isoformat())
    updated_messages=list(merge_message_runs(messages=[SystemMessage(content=TRUSTCALL_INSTRUCTION_FORMATTED)] + state["messages"][:-1]))

    # Initialize the spy for visibility into the tool calls made by Trustcall
    spy = Spy()
    
    # Create the Trustcall extractor for updating the ToDo list 
    todo_extractor = create_extractor(
    model,
    tools=[ToDo],
    tool_choice=tool_name,
    enable_inserts=True
    ).with_listeners(on_end=spy)

    # Invoke the extractor
    result = todo_extractor.invoke({"messages": updated_messages, 
                                    "existing": existing_memories})

    # Save the memories from Trustcall to the store
    for r, rmeta in zip(result["responses"], result["response_metadata"]):
        store.put(namespace,
                  rmeta.get("json_doc_id", str(uuid.uuid4())),
                  r.model_dump(mode="json"),
            )
        
    # Respond to the tool call made in task_mAIstro, confirming the update
    tool_calls = state['messages'][-1].tool_calls

    # Extract the changes made by Trustcall and add the the ToolMessage returned to task_mAIstro
    todo_update_msg = extract_tool_info(spy.called_tools, tool_name)
    return {"messages": [{"role": "tool", "content": todo_update_msg, "tool_call_id":tool_calls[0]['id']}]}

def update_instructions(state: MessagesState, config: RunnableConfig, store: BaseStore):

    """Reflect on the chat history and update the memory collection."""
    
    # Get the user ID from the config
    user_id = config["configurable"]["user_id"]
    
    namespace = ("instructions", user_id)

    existing_memory = store.get(namespace, "user_instructions")
        
    # Format the memory in the system prompt
    system_msg = CREATE_INSTRUCTIONS.format(current_instructions=existing_memory.value if existing_memory else None)
    new_memory = model.invoke([SystemMessage(content=system_msg)]+state['messages'][:-1] + [HumanMessage(content="Please update the instructions based on the conversation")])

    # Overwrite the existing memory in the store 
    key = "user_instructions"
    store.put(namespace, key, {"memory": new_memory.content})
    tool_calls = state['messages'][-1].tool_calls
    return {"messages": [{"role": "tool", "content": "updated instructions", "tool_call_id":tool_calls[0]['id']}]}

# Conditional edge
def route_message(state: MessagesState, config: RunnableConfig, store: BaseStore) -> Literal[END, "update_todos", "update_instructions", "update_profile"]:

    """Reflect on the memories and chat history to decide whether to update the memory collection."""
    message = state['messages'][-1]
    if len(message.tool_calls) ==0:
        return END
    else:
        tool_call = message.tool_calls[0]
        if tool_call['args']['update_type'] == "user":
            return "update_profile"
        elif tool_call['args']['update_type'] == "todo":
            return "update_todos"
        elif tool_call['args']['update_type'] == "instructions":
            return "update_instructions"
        else:
            raise ValueError

# Create the graph + all nodes
builder = StateGraph(MessagesState)

# Define the flow of the memory extraction process
builder.add_node(task_mAIstro)
builder.add_node(update_todos)
builder.add_node(update_profile)
builder.add_node(update_instructions)
builder.add_edge(START, "task_mAIstro")
builder.add_conditional_edges("task_mAIstro", route_message)
builder.add_edge("update_todos", "task_mAIstro")
builder.add_edge("update_profile", "task_mAIstro")
builder.add_edge("update_instructions", "task_mAIstro")

# Store for long-term (across-thread) memory
across_thread_memory = InMemoryStore()

# Checkpointer for short-term (within-thread) memory
within_thread_memory = MemorySaver()

# We compile the graph with the checkpointer and store
graph = builder.compile(checkpointer=within_thread_memory, store=across_thread_memory)

# View
display(Image(graph.get_graph(xray=1).draw_mermaid_png()))

# We supply a thread ID for short-term (within-thread) memory
# We supply a user ID for long-term (across-thread) memory 
config = {"configurable": {"thread_id": "1", "user_id": "Lance"}}

# User input to create a profile memory
input_messages = [HumanMessage(content="My name is Lance. I live in SF with my wife. I have a 1 year old daughter.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# User input for a ToDo
input_messages = [HumanMessage(content="My wife asked me to book swim lessons for the baby.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# User input to update instructions for creating ToDos
input_messages = [HumanMessage(content="When creating or updating ToDo items, include specific local businesses / vendors.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# Check for updated instructions
user_id = "Lance"

# Search 
for memory in across_thread_memory.search(("instructions", user_id)):
    print(memory.value)

# User input for a ToDo
input_messages = [HumanMessage(content="I need to fix the jammed electric Yale lock on the door.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# Namespace for the memory to save
user_id = "Lance"

# Search 
for memory in across_thread_memory.search(("todo", user_id)):
    print(memory.value)

# User input to update an existing ToDo
input_messages = [HumanMessage(content="For the swim lessons, I need to get that done by end of November.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# User input for a ToDo
input_messages = [HumanMessage(content="Need to call back City Toyota to schedule car service.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# Namespace for the memory to save
user_id = "Lance"

# Search 
for memory in across_thread_memory.search(("todo", user_id)):
    print(memory.value)

# We supply a thread ID for short-term (within-thread) memory
# We supply a user ID for long-term (across-thread) memory 
config = {"configurable": {"thread_id": "2", "user_id": "Lance"}}

# Chat with the chatbot
input_messages = [HumanMessage(content="I have 30 minutes, what tasks can I get done?")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()

# Chat with the chatbot
input_messages = [HumanMessage(content="Yes, give me some options to call for swim lessons.")]

# Run the graph
for chunk in graph.stream({"messages": input_messages}, config, stream_mode="values"):
    chunk["messages"][-1].pretty_print()