Getting Started with Mem0 and Milvus

Mem0 is an intelligent memory layer for AI applications, designed to deliver personalized and efficient interactions by retaining user preferences and continuously adapting over time. Ideal for chatbots and AI-driven tools, Mem0 creates seamless, context-aware experiences.

In this tutorial, we’ll cover essential Mem0 memory management operations—adding, retrieving, updating, searching, deleting, and tracking memory history—using Milvus, a high-performance, open-source vector database that powers efficient storage and retrieval. This hands-on introduction will guide you through foundational memory operations to help you build personalized AI interactions with Mem0 and Milvus.

Preparation

Download required libraries

$ pip install mem0ai pymilvus

If you are using Google Colab, to enable dependencies just installed, you may need to restart the runtime (click on the “Runtime” menu at the top of the screen, and select “Restart session” from the dropdown menu).

Configure Mem0 with Milvus

We will use OpenAI as the LLM in this example. You should prepare the api key OPENAI_API_KEY as an environment variable.

import os

os.environ["OPENAI_API_KEY"] = "sk-***********"

Now, we can configure Mem0 to use Milvus as the vector store

# Define Config
from mem0 import Memory

config = {
    "vector_store": {
        "provider": "milvus",
        "config": {
            "collection_name": "quickstart_mem0_with_milvus",
            "embedding_model_dims": "1536",
            "url": "./milvus.db",  # Use local vector database for demo purpose
        },
    },
    "version": "v1.1",
}

m = Memory.from_config(config)

Managing User Memories with Mem0 and Milvus

Adding a Memory

The add function stores unstructured text in Milvus as a memory, associating it with a specific user and optional metadata.

Here, we’re adding Alice’s memory, “working on improving my tennis skills,” along with relevant metadata for context to Milvus.

# Add a memory to user: Working on improving tennis skills
res = m.add(
    messages="I am working on improving my tennis skills.",
    user_id="alice",
    metadata={"category": "hobbies"},
)

res

{'results': [{'id': '77162018-663b-4dfa-88b1-4f029d6136ab',
   'memory': 'Working on improving tennis skills',
   'event': 'ADD'}],
 'relations': []}

Update a Memory

We can use the add function’s return value to retrieve the memory ID, allowing us to update this memory with new information via update.

# Get memory_id
memory_id = res["results"][0]["id"]

# Update this memory with new information: Likes to play tennis on weekends
m.update(memory_id=memory_id, data="Likes to play tennis on weekends")

{'message': 'Memory updated successfully!'}

Get All Memory For a User

We can use the get_all function to view all inserted memories or filter by user_id in Milvus.

Note that we can see the memory is now changed from “Working on impriving tennis skills” to "Likes to play tennis on weekends".

# Get all memory for the user Alice
m.get_all(user_id="alice")

{'results': [{'id': '77162018-663b-4dfa-88b1-4f029d6136ab',
   'memory': 'Likes to play tennis on weekends',
   'hash': '4c3bc9f87b78418f19df6407bc86e006',
   'metadata': None,
   'created_at': '2024-11-01T19:33:44.116920-07:00',
   'updated_at': '2024-11-01T19:33:47.619857-07:00',
   'user_id': 'alice'}]}

View Memory Update History

We can also view the memory update history by specifying which memory_id we are interested in via history function.

m.history(memory_id=memory_id)

[{'id': '71ed3cec-5d9a-4fa6-a009-59802450c0b9',
  'memory_id': '77162018-663b-4dfa-88b1-4f029d6136ab',
  'old_memory': None,
  'new_memory': 'Working on improving tennis skills',
  'event': 'ADD',
  'created_at': '2024-11-01T19:33:44.116920-07:00',
  'updated_at': None},
 {'id': 'db2b003c-ffb7-42e4-bd8a-b9cf56a02bb9',
  'memory_id': '77162018-663b-4dfa-88b1-4f029d6136ab',
  'old_memory': 'Working on improving tennis skills',
  'new_memory': 'Likes to play tennis on weekends',
  'event': 'UPDATE',
  'created_at': '2024-11-01T19:33:44.116920-07:00',
  'updated_at': '2024-11-01T19:33:47.619857-07:00'}]

Search Memory

We can use search function to look for the most related memory for the user.

Let’s start by adding another memory for Alice.

new_mem = m.add(
    "I have a linear algebra midterm exam on November 20",
    user_id="alice",
    metadata={"category": "task"},
)

Now, we call get_all specifying the user_id to verify that we have indeed 2 memory entries for user Alice.

m.get_all(user_id="alice")

{'results': [{'id': '77162018-663b-4dfa-88b1-4f029d6136ab',
   'memory': 'Likes to play tennis on weekends',
   'hash': '4c3bc9f87b78418f19df6407bc86e006',
   'metadata': None,
   'created_at': '2024-11-01T19:33:44.116920-07:00',
   'updated_at': '2024-11-01T19:33:47.619857-07:00',
   'user_id': 'alice'},
  {'id': 'aa8eaa38-74d6-4b58-8207-b881d6d93d02',
   'memory': 'Has a linear algebra midterm exam on November 20',
   'hash': '575182f46965111ca0a8279c44920ea2',
   'metadata': {'category': 'task'},
   'created_at': '2024-11-01T19:33:57.271657-07:00',
   'updated_at': None,
   'user_id': 'alice'}]}

We can perform search now by providing query and user_id. Note that we are by default using L2 metric for similarity search, so a smaller score means greater similarity.

m.search(query="What are Alice's hobbies", user_id="alice")

{'results': [{'id': '77162018-663b-4dfa-88b1-4f029d6136ab',
   'memory': 'Likes to play tennis on weekends',
   'hash': '4c3bc9f87b78418f19df6407bc86e006',
   'metadata': None,
   'score': 1.2807445526123047,
   'created_at': '2024-11-01T19:33:44.116920-07:00',
   'updated_at': '2024-11-01T19:33:47.619857-07:00',
   'user_id': 'alice'},
  {'id': 'aa8eaa38-74d6-4b58-8207-b881d6d93d02',
   'memory': 'Has a linear algebra midterm exam on November 20',
   'hash': '575182f46965111ca0a8279c44920ea2',
   'metadata': {'category': 'task'},
   'score': 1.728922724723816,
   'created_at': '2024-11-01T19:33:57.271657-07:00',
   'updated_at': None,
   'user_id': 'alice'}]}

Delete Memory

We can also delete a memory by providing the corresponding memory_id.

We will delete the memory “Likes to play tennis on weekends” as its memory_id has already been retrieved, and call get_all to verify the deletion is successful.

m.delete(memory_id=memory_id)

m.get_all("alice")

{'results': [{'id': 'aa8eaa38-74d6-4b58-8207-b881d6d93d02',
   'memory': 'Has a linear algebra midterm exam on November 20',
   'hash': '575182f46965111ca0a8279c44920ea2',
   'metadata': {'category': 'task'},
   'created_at': '2024-11-01T19:33:57.271657-07:00',
   'updated_at': None,
   'user_id': 'alice'}]}