langchain-recall

Recall memory for LangChain and LangGraph agents. Recall learns about the people your agent talks to from their conversations, and gives that back as context, search results and answers.

Install

pip install "langchain-recall[agents]"    # without [agents] if you do not use create_agent

Create an API key at https://app.recallmem.dev/keys and export it as RECALL_API_KEY (and RECALL_WORKSPACE_ID to pick a workspace). Running your own server? Set RECALL_URL=http://localhost:8000.

Quickstart

from langchain.agents import create_agent
from langchain_recall import RecallMiddleware, RecallToolkit

agent = create_agent(
    model,  # any LangChain chat model
    tools=RecallToolkit().get_tools(),  # recall_search, recall_ask, recall_remember
    middleware=[RecallMiddleware()],  # memory in the prompt, every exchange recorded
)
agent.invoke(
    {"messages": [{"role": "user", "content": "Plan my weekend, you know what I like."}]},
    {"configurable": {"thread_id": "chat-1"}},
    context={"user_id": "user-42"},
)

The middleware adds the user's peer card and the facts Recall learned that relate to the new message to the system prompt (no model call, well under a second), then stores the user's message and the final answer in the Recall session named after the thread. Tool calls and their results are not recorded. The user comes from the run's context, context={"user_id": ...}, or pass user_id= to a component when there is one user.

What is in the package

Every component has blocking and async methods. Pass client=Recall(...) (serves both) or AsyncRecall(...) (async only) from recall-mem; without one they use Recall(), configured from the environment.

from langchain_core.messages import AIMessage, HumanMessage
from langchain_recall import RecallChatMessageHistory, RecallRetriever

history = RecallChatMessageHistory("chat-1", user_id="user-42")
history.add_messages([HumanMessage("I moved to Utrecht."), AIMessage("Welcome to Utrecht!")])
docs = RecallRetriever(user_id="user-42", source="facts").invoke("where does the user live?")

RecallChatMessageHistory stores human messages as the user's (Recall learns from them) and AI messages as the assistant's (read as context, not learned about). Tool and system messages are kept for the history only. Every message comes back exactly as it was added. clear() hides the messages from the history; Recall keeps them and what it learned. Delete the Recall session to erase them. LangChain marks BaseChatMessageHistory as deprecated in favour of LangGraph persistence; for new agents use the middleware.

LangGraph. The tools and the retriever work in any graph node or ToolNode, and read the user from the run's context when you leave user_id out. There is no Recall BaseStore or checkpointer: a store is a key-value document store and a checkpointer saves graph state, and Recall is neither. It learns from conversations. Keep your checkpointer for graph state and use Recall for what the agent should remember.

Docs

• Guide: https://recallmem.dev/guides/langchain
• Recall docs: https://recallmem.dev/docs

MIT licensed.