에이전트에 메모리와 대화 상태 부여하기

LLMs are stateless: each call knows nothing about the last unless you tell it. Without memory, an agent forgets your name the instant you say it.

Memory ultimately means stuffing prior info into the context window. That window is finite, so the real challenge is deciding what to keep and what to drop.

Buffer memory stores the whole conversation and replays it every turn. Accurate, but it grows without bound and eventually overflows the context window.

Windowed memory keeps only the last N exchanges. It bounds size and forgets older context — perfect when only recent turns matter.

Summary memory periodically condenses older turns with the LLM, keeping the gist plus recent messages — preserving meaning in a small footprint.

Two flavors serve different needs: short-term memory holds the current chat in the prompt, while long-term persists facts across sessions in a store.

For knowledge that must survive sessions, store messages as embeddings in a vector store and retrieve the most relevant ones by similarity instead of replaying everything.

Wire memory into a conversation chain. Each call loads prior context, runs the model, and saves the new exchange automatically.

Real apps serve many users at once. Scope memory by session or user id so separate conversations never leak into each other.

More memory means more tokens — higher cost and latency. Long-term stores may hold sensitive data, so mind retention limits and what you're allowed to keep.

Choosing a strategy: buffer or window for short chats, summary for long ones, vector for cross-session facts — always scoped per user and mindful of cost.

You've met several memory types — which fits when? Time to put it to the test.

Recap: memory feeds prior context into a finite window. Buffer, window, and summary trade accuracy for size, vector stores enable long-term recall — always scope and watch cost.