หน่วยความจำและบริบทการสนทนาใน LangChain

A language model has no memory between calls. Each request is independent — it only knows what you put in the current prompt. To build a chatbot that remembers, you must feed prior turns back in yourself.

LangChain memory automates this: it stores the conversation and injects relevant history into the prompt on each new turn. Your chain stays simple while the model appears to remember.

The simplest memory keeps the full transcript and prepends it to every prompt. Great for short chats, but it grows with every turn.

Every model has a finite context window (a token budget). A growing buffer eventually overflows it, causing errors or truncation. Managing history size is the central challenge of memory.

ConversationBufferWindowMemory keeps only the last k turns. It bounds token usage at the cost of forgetting older context — a simple, effective trade-off for many chatbots.

ConversationSummaryMemory uses the LLM to compress old turns into a running summary. You keep the gist of a long conversation in far fewer tokens, sacrificing exact wording for breadth.

ConversationSummaryBufferMemory keeps recent turns verbatim and summarizes everything older. Recent context stays precise while distant context is condensed — a popular default for production chatbots.

You attach memory to a conversational chain. On each call, the chain loads history, builds the prompt, calls the model, and saves the new turn back to memory automatically.

Memory exposes its content as a variable (often history or chat_history) that your prompt template references. The placeholder is where the stored conversation gets injected.

In-process memory vanishes when the app restarts. For real users, back memory with a store — Redis, a database, or a chat-message-history backend keyed by session id — so conversations survive across requests and servers.

Match memory to need: buffer for short chats, window when you only care about recent turns, summary for long sessions on a budget, and summary-buffer for the common case. Always persist memory for multi-user apps.

Test your understanding of LangChain memory.

You learned how LangChain gives chatbots memory:

• LLMs are stateless; memory re-injects history each turn
• Buffer, window, summary, and summary-buffer trade detail against tokens
• Memory exposes a history variable that the prompt template uses
• Persist memory per session for multi-user, multi-server apps