Mem0

Mem0 is a long-term memory layer for AI agents and LLM applications. Mem0 extracts meaningful facts about the user from conversations, stores them as embeddings, and on subsequent queries finds relevant memories by semantic similarity. This allows the agent to "remember" the user's preferences and context across sessions.

Integration with YDB allows using YDB as a vector store for Mem0 memory — without a separate specialized vector database. The integration works on top of LangChain: Mem0 supports LangChain vector stores as a backend, and YDB provides a compatible store via the langchain-ydb package.

Installation

To use this integration, install a local YDB. For more information, see Install and start YDB.

Install Mem0 and the YDB integration package with LangChain:

pip install -qU mem0ai langchain-ydb langchain-openai

Initialization

Create a YDB vector store using langchain-ydb and pass it to the Mem0 configuration with the langchain provider:

The key to OpenAI models is expected in the OPENAI_API_KEY environment variable — set it in advance in your environment or secret manager.

from langchain_openai import OpenAIEmbeddings
from langchain_ydb.vectorstores import YDB, YDBSearchStrategy, YDBSettings
from mem0 import Memory

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

vector_store = YDB(
    embeddings,
    config=YDBSettings(
        host="localhost",
        port=2136,
        database="/local",
        table="mem0_memories",
        strategy=YDBSearchStrategy.COSINE_SIMILARITY,
    ),
)

memory = Memory.from_config({
    "vector_store": {
        "provider": "langchain",
        "config": {"client": vector_store},
    },
    "embedder": {
        "provider": "langchain",
        "config": {"model": embeddings},
    },
})

Saving facts

Pass the conversation messages to Mem0 — the model will extract facts from them and save them to the YDB table. Memory is shared between users using user_id:

memory.add(
    [
        {"role": "user", "content": "Привет! Меня зовут Алиса, я работаю архитектором."},
        {"role": "assistant", "content": "Рада знакомству, Алиса!"},
        {"role": "user", "content": "Предпочитаю примеры на Python."},
    ],
    user_id="alice",
)

Memory search

Before forming a response, the agent searches for relevant memories — Mem0 performs a vector search in YDB and returns facts ordered by similarity:

results = memory.search("На каком языке показывать примеры?", user_id="alice")
for item in results["results"]:
    print(f"{item['score']:.3f}  {item['memory']}")

Result:

0.512  Prefers examples in Python
0.231  Works as an architect

Viewing and deletion

User memories can be retrieved or deleted entirely:

# all user memories
for item in memory.get_all(user_id="alice")["results"]:
    print(item["memory"])

# delete all user memories
memory.delete_all(user_id="alice")