Build Windowed Memory Yourself

A plain ConversationBufferMemory grows with every single message, with no upper bound. In a long-running conversation — a customer support session spanning dozens of exchanges — that buffer eventually exceeds the model's context window entirely, causing either a hard failure or silent truncation you don't control.

Windowed memory solves the size problem by genuinely discarding old messages — this lesson's quiz question isn't rhetorical. Ask about something from outside the window, and the model has no way to know, because the fact simply isn't in the prompt anymore. This is the same fundamental trade-off as the rolling window truncation technique from earlier context-management concepts, applied specifically to LangChain's memory abstraction.

ConversationBufferMemory keeps everything, forever — fine for a short chat, a real problem for a long-running one, since eventually the buffer exceeds the model's context window. LangChain's ConversationBufferWindowMemory fixes this by keeping only the last k exchanges.

Build Windowed Memory Yourself. With k=2, only the last 2 user/AI exchanges (4 messages total) should survive. Finish get_buffer_string(): slice self.messages down to the last (self.k * 2) messages before formatting — the earliest exchange should be dropped entirely.

With this k=2 windowed memory, if you now asked 'What's my name?', what would happen?

• →The model would not know — the 'My name is Sam' exchange has fallen outside the k=2 window and is no longer part of the injected context, so the model has genuinely lost access to that fact.
• →The model would still answer correctly, because memory never truly forgets anything.

Module 3 complete: you've built full-history and windowed memory, and proven memory works on a real model. Module 4 gives your chatbot access to external documents — the actual RAG mechanics, built LangChain-style this time.