Memory.from_config: all three pieces

Two stand-ins exist and neither is plugged in. The configuration is one dictionary with one entry per piece, and one detail in it is not obvious.

The detail

Mem0 picks a provider by name from a fixed list. You cannot add "pretend" to that list from outside: the validator raises Unsupported embedding provider. The supported way to pass your own object is the langchain provider, which takes an instance instead of a name.

"llm": {"provider": "langchain", "config": {"model": PretendModel()}},
"embedder": {"provider": "langchain", "config": {"model": PretendEmbedder()}},

That is why lesson 3 installed the langchain packages. They are not Mem0's dependency; they are the door your own code comes through.

The store

"vector_store": {"provider": "qdrant", "config": {
    "path": tempfile.mkdtemp(prefix="mem0-"),
    "embedding_model_dims": 64,
    "on_disk": False,
}},

Qdrant runs inside the process, so there is no server and no Docker. path is where it keeps its files and embedding_model_dims has to match the embedder from lesson 6. A fresh temporary directory each time means every lesson in this course starts empty, which is what makes their output the same every run.

The helper

Put together, that is a dictionary long enough to be worth writing once. This goes at the bottom of the same file as the two stand-ins.

def memory(**extra):
    """A Mem0 memory that uses the stand-ins and a fresh directory each time."""
    from mem0 import Memory

    config = {
        "llm": {"provider": "langchain", "config": {"model": PretendModel()}},
        "embedder": {"provider": "langchain", "config": {"model": PretendEmbedder()}},
        "vector_store": {"provider": "qdrant", "config": {
            "path": tempfile.mkdtemp(prefix="mem0-"),
            "embedding_model_dims": 64,
            "on_disk": False,
        }},
    }

    config.update(extra)
    return Memory.from_config(config)

extra lets a later lesson override one piece without repeating the rest, which lessons 17 and 18 both use.

All of it running

from pretend_mem0 import memory

shop = memory()
result = shop.add("I prefer email updates, not SMS. Deliver to my office.",
                  user_id="ravi")
for stored in result["results"]:
    print(stored["event"], "|", stored["memory"])

Two memories out of one sentence, with no key and no network. Everything from here is what you can do with them.