Question 1

How does AI memory work in n8n and why does the session key matter?

Accepted Answer

State the mechanism, then frame the session key as a security matter — that is what sets the answer apart. "Memory attaches to an agent node as a sub-node and keeps the conversation's earlier messages. The simplest in n8n is Simple Memory, which stores a configurable number of messages for the current session. For durability there are options backed by external services, such as Postgres Chat Memory and Redis Chat Memory. An important limitation: memory works only with agents; chains do not support it. The session key determines whose memory it is. There are two mistakes and both are serious: A fixed key makes every user share one memory — one user sees another's conversation. That is not a glitch, it is a data leak. A key that changes per message (the message ID, say) means memory effectively does nothing. The right key is a stable value identifying the user: the chat ID in Telegram, the user ID in Slack."

Question 2

How do you control the cost of memory in long conversations?

Accepted Answer

Explain the mechanism, then give the levers. "The root of the problem: every message held in memory is resent to the model on the next request. So the longer the conversation, the more each question costs. The 50th question in a 50-message conversation effectively pays for 49 messages' tokens. I use three levers: 1. Window size. I match the memory limit to what the conversation actually needs. In a support bot the last few messages are usually enough; the full history is not. 2. Summarisation. In long conversations the older part can be replaced by a summary rather than kept verbatim. n8n's Chat Memory Manager node exists for exactly this — checking the memory size and reducing it. 3. Ending the session. I define when a conversation is over. When a support ticket closes, the memory should close too, or the next question inherits stale context. One addition: you have to reduce everything that reaches the model, not just memory. Data returned by tools also lands in the context and is often more expensive than the history."

Question 3

How do you test an AI agent while building it?

Accepted Answer

Describe the working method — that is exactly what this question probes. "I start with the Chat Trigger. It opens a chat window inside n8n, so I see how the agent behaves without setting up Telegram or Slack. That is the fastest route on day one. I add tools one at a time. Attaching five at once and asking "why is it wrong?" is pointless. I connect one, confirm it gets chosen correctly, then add the next. I read the execution history. An agent runs several times per execution and every step is recorded: which tool it called, what came back, what it did next. It is the only way to see the agent's reasoning. I try the edge cases. Not just the questions I expect: an unknown customer, an empty result, a nonsense question, a very long input. Does the agent invent what it cannot find? I measure the result. n8n's evaluation features score output against known test cases. That is the only way to know whether a prompt or model change lowered quality — rather than a subjective "looked fine to me"."

Memory option	Where it is stored	When to choose it
Simple Memory	In the n8n instance's memory, for the current session	Getting started, testing, short conversations
Postgres Chat Memory	In a Postgres database	When the conversation must survive a restart
Redis Chat Memory	In Redis	When speed matters and there are many concurrent sessions
Chat Memory Manager	An extra node that manages memory	For controlling memory size or injecting synthetic messages

Chat Trigger and memory

📚 Sources and documentation