How do you use LangChain to create a conversational chain?

What you need to know

The model itself remembers nothing between calls. "Memory" means your code sends the earlier messages again with each new one.

1from langchain_core.messages import SystemMessage, trim_messages2from langchain_core.messages.utils import count_tokens_approximately3from langgraph.graph import StateGraph, MessagesState, START4from langgraph.checkpoint.memory import InMemorySaver56SYSTEM = SystemMessage("You are PayEase support. Answer from the help centre only.")78def respond(state: MessagesState):9    recent = trim_messages(state["messages"], strategy="last", max_tokens=3000,10                           token_counter=count_tokens_approximately, start_on="human")11    return {"messages": [llm.invoke([SYSTEM, *recent])]}1213builder = StateGraph(MessagesState)14builder.add_node("respond", respond)15builder.add_edge(START, "respond")16bot = builder.compile(checkpointer=InMemorySaver())1718cfg = {"configurable": {"thread_id": "user-812"}}19bot.invoke({"messages": [{"role": "user", "content": "My salary credit failed"}]}, cfg)20bot.invoke({"messages": [{"role": "user", "content": "Which bank was it sent to?"}]}, cfg)

• MessagesState holds a messages list; returning new messages appends them.
• The checkpointer saves the list after each turn; the same thread_id loads it back.
• trim_messages keeps only the most recent messages that fit a token budget, starting on a human turn.
• If you also need tools, create_agent(model, tools, checkpointer=...) gives the same persistence with less code.

Where history is stored

The legacy APIs

ConversationChain with ConversationBufferMemory stored history inside the chain object. RunnableWithMessageHistory wrapped a chain with a get_session_history function; it still runs but now emits a deprecation warning in langchain-core 1.6, and so does InMemoryChatMessageHistory. Know them for older codebases, but build new work on LangGraph persistence.

A real-life example

A payroll company's support bot over its help-centre docs stored chat history in a Python dict inside the web process. Each deploy wiped every conversation, and with four servers behind a load balancer, a user's second message often landed on a server that had never seen the first. Moving to a LangGraph graph with PostgresSaver, keyed by thread_id = chat session ID, fixed both problems. Adding trim_messages with a 3,000-token budget kept long chats, some over 60 turns, from failing with context-length errors, and cut the average prompt size by about 40%.

Follow-up questions to expect

• "How do you keep facts from early in a long chat?" — Summarise older turns into a short system note (LangChain's SummarizationMiddleware does this for agents), or save key facts to a long-term store.
• "How do you separate users?" — One thread_id per conversation, and check on the server that the thread belongs to the logged-in user.
• "How do you clear a conversation?" — Start a new thread_id, or delete that thread's checkpoints.