An ops lead wires up n8n's AI Agent node for WhatsApp support, drags in whatever memory node the node picker suggests first, and ships it on a Friday. By Monday there are two very different complaints on the table: either the agent is asking for an order number it was already given twice in the same thread, or finance is asking why the OpenAI bill is four times normal. Same root cause, both times: the memory node was never actually chosen, it was left on a default nobody looked at twice.
What is n8n actually storing when you attach a 'memory' node?
None of n8n's memory options give the underlying model real persistent knowledge between calls, the model itself has no memory of anything outside the current API request. What a memory node does is decide what gets re-injected into the prompt on the next call: n8n's AI Agent node needs a memory sub-node wired in, and the options behave very differently, Simple Memory (a fixed sliding window of the last N messages), full conversation/buffer memory (the entire transcript resent every call), Vector Store memory (semantically relevant past turns retrieved via embeddings), and persistent chat memory backed by Postgres or Redis (durable across a workflow restart).
Which one you pick determines exactly what the model sees on the next turn, and how much you pay to send it. Get that choice wrong and you land on one of two failure modes: a forgetful agent, or an expensive one.
Why the 'safe-looking' default becomes a cost problem
Full conversation memory means every single call to the agent resends the whole transcript built up so far, that's the definition of the option, not a bug. It's invisible on message two of a thread. By message thirty in a long support conversation, every new reply is re-paying, in tokens, for messages one through twenty-nine all over again, on top of the new one.
n8n's own guide to AI agent memory puts it plainly: "passing the full interaction history on every call burns through the budget fast." This is the same token-cost mechanic covered in reducing AI API costs in production, applied specifically to the memory layer instead of the model choice.
Window vs Buffer vs Vector Store vs persistent: what each is actually for
| Memory type | What it re-injects into the prompt | Grows with conversation length? | Best fit |
|---|---|---|---|
| Simple / Window Memory | Only the last N messages | No, fixed size | Short transactional threads: order status, FAQ triage, routing |
| Full conversation / Buffer Memory | The entire transcript, every single turn | Yes, linearly | Short-lived, low-volume chats where nothing can be missed |
| Vector Store Memory | Semantically similar past turns or documents, retrieved via embeddings | No, scales independent of history length | Long-running or repeat-customer agents that need to recall something from weeks ago |
| Persistent (Postgres / Redis) Memory | Conversation history that survives a workflow restart | Usually paired with Window Memory | Anything where a crash or restart mid-conversation can't be allowed to erase context |
Source: n8n Blog — AI Agent Memory: Types, Storage, and Retrieval Guide.
The setup errors that break the AI Agent node before memory is even the issue
- Missing Chat Model connection: the node throws an error if it's executed without a Chat Model attached, click the "+ Chat Model" button on the node to fix it, per n8n's AI Agent node documentation.
- Simple Memory sub-node errors: usually caused by an outdated version of the node, removing it and re-adding it from scratch pulls the current version and clears the error.
- Null value in the Prompt input: happens when an expression in the Text field resolves to null, commonly an empty chatInput field when the workflow uses a connected Chat Trigger node.
- "No prompt specified": occurs when the Prompt setting is left on "Connected Chat Trigger Node" but nothing upstream is actually feeding it, switching it to "Define below" lets you construct the prompt manually from data references.
A simple rule for picking memory without overthinking it
For a short, bounded conversation, an FAQ bot, a routing agent, an order-status lookup, Simple/Window Memory set to 5-10 messages is both the cheapest option and functionally enough, per the practitioner guidance in n8n's own memory documentation.
For an agent that needs to recall something a customer mentioned weeks ago, the fix isn't a bigger window, it's adding Vector Store memory alongside Window memory, not instead of it. That hybrid, a small working-memory window for immediate context plus a vector store for long-term recall, is the pattern most production agents actually run. If a crash or redeploy mid-conversation would be unacceptable, for example a paid support agent mid-escalation, back that memory with Postgres or Redis Chat Memory so a restart doesn't erase the thread.
How AIBOOTSTRAPPER solved this for AI Doctor
AI Doctor is AIBOOTSTRAPPER's bilingual (Arabic and English) symptom-triage assistant, engineered with safety guardrails and a handoff workflow that escalates to a human doctor with a complete, pre-filled summary, cutting consultation prep time by 68%. A triage conversation has to hold onto exactly what a patient already said across several turns, without re-paying for the entire session on every follow-up question and without losing that context if a session gets interrupted, which is precisely the memory-selection discipline this article walks through, applied to a case where getting it wrong has real consequences.
How AIBOOTSTRAPPER helps
AIBOOTSTRAPPER's AI automation team builds n8n and custom AI agent workflows with the memory architecture chosen deliberately for what the agent actually needs to remember, not left on whatever the node picker suggests first, so a client's agent doesn't quietly become forgetful or expensive as usage grows.
If you're not sure which memory type your current n8n agent is even running on, book a call and we'll audit the workflow before it becomes a bigger bill or a customer complaint.
Want this done for you?
Book a free strategy call and we'll show you how to build and market your business with AI.