Mem0 integration with Crew ai

General
,
5

The Problem and Some Theory

Tracing the error, I noticed that when running a Task, the Agent tries to build context using ContextualMemory. This involves fetching user context via _fetch_user_context, which in turn calls self.um.search(query). However, self.um (which should be an instance of UserMemory) is None, leading to the AttributeError: 'NoneType' object has no attribute 'search' error we’re seeing.

The problem originates during the initialization of the Crew class. When the create_crew_memory method is called, it does the following:

  • It checks self.memory. Since we’re passing the parameter memory=True, this evaluates to True.
  • It then initializes _long_term_memory, _short_term_memory, and _entity_memory. From what I’ve seen, this part works correctly and doesn’t affect our issue, so let’s move on.
  • Crucially, it then checks: if self.memory_config and "user_memory" in self.memory_config, and here lies the bug. If the condition is NOT met (which is the case with the Mem0 example config), it explicitly sets self._user_memory = None. This is the direct cause of self.um being None later on.

After this point, a ContextualMemory object is created with: self.crew._short_term_memory (:grinning_face:), self.crew._long_term_memory (:smiley:), self.crew._entity_memory (:grinning_face_with_smiling_eyes:), and self.crew._user_memory (:pensive_face:).

And that, in a nutshell, is how self.um becomes None, causing the self.um.search(query) method to throw the error (goes without saying this should probably be wrapped in a try block, right?).

A Workaround for Mem0 (Includes Functional Code)

As you saw above, we need to force the check if self.memory_config and "user_memory" in self.memory_config to evaluate to True. Therefore, we need to ensure the user_memory key is present in the dictionary passed to the memory_config parameter.

Here’s the functional code. Hope it helps you guys out.

from crewai import Agent, Task, Crew, LLM, Process
from mem0 import MemoryClient
import os

os.environ['MEM0_API_KEY'] = ''
os.environ['GEMINI_API_KEY'] = ''

# Step 1: Record preferences based on user input

client = MemoryClient()
messages = [
    {
        'role': 'user', 
        'content': 'Hi! I\'m planning a vacation and could use some advice.'
    },
    {
        'role': 'assistant', 
        'content': 'Hello! I\'d be happy to help with your vacation planning. '
                   'What kind of destination do you prefer?'
    },
    {
        'role': 'user', 
        'content': 'I am more of a beach person than a mountain person.'
    },
    {
        'role': 'assistant', 
        'content': 'That\'s interesting. Do you like hotels or Airbnb?'
    },
    {
        'role': 'user', 
        'content': 'I like Airbnb more.'
    }
]
client.add(messages, user_id='John Doe', output_format='v1.1')

# Step 2: Create a Crew with User Memory

gemini_llm = LLM(
    model='gemini/gemini-2.0-flash',
    temperature=0.7,
	max_tokens=1024
)

travel_planner_agent = Agent(
    role='Travel Planner',
    goal='Create a detailed travel itinerary',
    backstory='Expert in travel logistics',
    llm=gemini_llm,
    verbose=True,
    allow_delegation=False
)

planning_task = Task(
    description=(
        'Plan a weekend trip to {destination} focusing on beaches, '
        'considering the user preference for Airbnb.'
    ),
    expected_output=(
        'A day-by-day itinerary in markdown format, including beach '
        'and activity suggestions.'
    ),
    agent=travel_planner_agent
)

# --- Workaround Explanation ---
#
# The Crew class currently has a bug where it doesn't automatically
# create the UserMemory component when 'provider': 'mem0' is set.
# It incorrectly looks for a 'user_memory' key in the config.
# To trigger the correct initialization path *within* UserMemory,
# we add the 'user_memory': {} key manually. This satisfies the
# faulty check in Crew, causing it to call UserMemory(crew=self),
# which then correctly uses the 'provider' and 'config' details.

crew = Crew(
    agents=[travel_planner_agent],
    tasks=[planning_task],
    verbose=True,
    process=Process.sequential,
    memory=True,
    memory_config={
        'provider': 'mem0',
        'config': {
            'user_id': 'John Doe'
        },
        'user_memory': {}   # Workaround for buggy memory initialization
    },
)

result = crew.kickoff(
    inputs={
        'destination': 'Cancun'
    }
)

print(f'\n🤖 Your Travel Plan:\n\n{result.raw}\n')
1 Like