A chat that remembers the conversation

Remembers what was said earlier, so the customer does not have to repeat themselves every message.

What the scenario looks like

Сообщение клиента

Ответ с учётом переписки▶▦

Отправка ответа▶

Модель▦

Память диалога▶S

Execute + Data
LLM
Memory

Open in your workspace →

A bot with no memory makes people repeat themselves: “I already said the order is for Kazan.” That is the most visible difference between answering and having a conversation, and it costs one node.

What is inside

Customer message — the entry.
Conversation memory — a window_memory node wired to the agent with link_memory. It keeps the recent turns of this conversation and hands them to the model along with the new question.
Answer in context — the agent.
Send the answer — the exit.

Memory fills up from this scenario’s own earlier runs. On the first message it is empty, exactly as it is for a bot you just published. And it remembers the conversation, not your catalogue: for the bot to know your delivery terms you need a knowledge base.

What to set up for yourself

Window depth (window_size): how many recent turns go to the model. Every turn in the window is paid for in input tokens on every subsequent question, so a hundred-message window is not a smarter bot, it is a more expensive turn.
Memory scope (scope). thread means one conversation, one memory: every visitor and every messenger chat gets their own. That is what you want in most cases.
The agent’s system prompt: who your bot is and what it talks about.

What to connect

Nothing. Memory is stored on the platform and tied to the conversation; there is no storage to set up. If what you want is memory about the person rather than the conversation — “customer from Kazan, orders monthly” — that is a different node, agent_memory, and a different scenario.