n8n AI BuilderPick your trigger, model, memory, and tools. Get the exact node graph.

n8n's AI Agent node sits in the middle of a workflow and makes one decision over and over: given the conversation so far, does it answer directly, or call a tool first. You feed it a chat model as a required sub-node input, and that model is what reasons about the decision. Everything else, the trigger, the tools, the memory, is plumbing around that one node.

By default the AI Agent runs as a Tools Agent. It uses the connected model's native function-calling API, which is the setting n8n recommends whenever your model supports it. There is a second mode, Conversational Agent, that describes the available tools inside the system prompt and parses the model's text output for a tool call instead of using a structured API. That fallback exists for models that never got function calling, or for simpler chat interactions where a full agent loop is overkill.

The practical implication: pick Tools Agent unless you have a specific reason not to. It is more reliable at actually calling the tool you meant, and it is what most current chat-model APIs (OpenAI, Anthropic, Gemini, and the tool-calling-capable Ollama models) are built to work with.

Memory in n8n is a sub-node you attach to the AI Agent, not a setting you toggle on. Skip it and the agent has no idea what was said two messages ago, which is fine for a single-shot tool and bad for anything conversational.

Simple Memory keeps the conversation in n8n's own running process. It works, and it is the fastest way to test a multi-turn agent, but the history does not survive the workflow finishing or n8n restarting. If your chat widget is expected to remember a user across page reloads or your server redeploys, Simple Memory will quietly lose that context.

Postgres Chat Memory and Redis Chat Memory write history to an actual database keyed by a session ID, so conversations survive restarts and scale past whatever fits in n8n's process memory. For a webhook-triggered agent, that session ID usually comes from the caller (a user ID, a ticket number) rather than being generated by n8n itself, so plan for that field when you design the trigger payload.

Tools are what turn a chatbot into something that does work. Each tool you attach under the AI Agent node becomes a function the model can choose to call, with a description it uses to decide when that call makes sense.

The four tool types in the builder above cover most real workflows. HTTP Request Tool lets the agent call any REST API you point it at. Code Tool runs a JavaScript or Python snippet you write, useful for calculations or formatting the model should not be trusted to do free-form. Vector Store Tool gives the agent retrieval over your own documents, but it needs an embeddings node and a vector store node (Pinecone, Qdrant, Supabase Vector Store, and similar are all supported) wired underneath it to actually have something to search. Call n8n Workflow Tool lets one agent trigger a separate n8n workflow as a sub-task, which is how you keep a single agent from turning into one enormous, unmaintainable canvas.

This page is deliberately n8n-only, because the node-level detail above does not generalize across platforms. The short version: n8n wins on self-hosting, per-execution pricing that does not scale per task, and control over exactly what each node does. Zapier AI Agents win on breadth, a prompt-driven builder, and not having to run any infrastructure yourself. Make sits in between with more branching logic than Zapier but a smaller AI-specific node set than n8n.

If you are still choosing a platform rather than committing to n8n, the comparison and 6-way picker linked above walks through triggers, budget, and code comfort in more detail than fits here.