Real models: model strings and keys

Anywhere LangMem takes a model, a string like anthropic:claude-sonnet-4-5 works too. LangChain turns it into a chat model that reads its key from the environment.

Example

manager = create_memory_manager("anthropic:claude-sonnet-4-5")

A string is passed to LangChain's init_chat_model. The provider's integration package must be installed, and its key set:

Model string	Package	Key
`anthropic:claude-sonnet-4-5`	`langchain-anthropic`	`ANTHROPIC_API_KEY`
`openai:gpt-4.1-mini`	`langchain-openai`	`OPENAI_API_KEY`
`ollama:llama3.2`	`langchain-ollama`	none; runs on your machine

Example

from langmem import create_memory_manager

manager = create_memory_manager("anthropic:claude-sonnet-4-5")
manager.invoke({"messages": [{"role": "user", "content": "Please email me."}]})

With no key, the call fails before anything is sent, and the message names the variable to set. The rest of the course passes MemoryModel() where this passes a string.

Memory needs a capable model

Extraction works through parallel tool calls with nested JSON Patch arguments. Small local models often get those wrong; a model that is good at tool calling matters more here than a model that writes well.

Try it yourself

Pass "openai:gpt-4.1-mini" and read the error.
Create the model yourself with init_chat_model("anthropic:claude-sonnet-4-5", temperature=0) and pass the object.
List what a model needs to do for LangMem, from lesson 3's request.

PreviousMemoryModel: a rule-based chat model

Every expert started right here.