Practical Patterns for Capturing the Last Message from an n8n AI Agent (Gemini)

In n8n workflows the simplest point to capture an AI agent's last message is immediately after the node that invokes the model: the node's execution output contains the response payload you can map into a variable via a Set node or Function node. For users of hosted Gemini-like nodes, ensure you inspect the node's raw output field (often labeled "response" or "choices") and map the intended string to a named workflow variable right away to avoid losing context in downstream transformations.

For more robust patterns, persist the last message to a lightweight datastore (Redis, DynamoDB, or a simple key-value table) keyed by conversation ID. This is essential for multi-step agents or retries: it preserves state across executions, supports deduplication, and enables retrieval even if the workflow is resumed or an execution fails. Use an idempotent write pattern (write-if-not-exists or versioned updates) to prevent race conditions when multiple triggers interact with the same conversation.

Operational concerns matter: validate and sanitize model outputs before reuse, log both request and response for auditability, and enforce retention policies for PII. If you need message-level metadata (timestamps, tokens used, model version), store those alongside the text to enable cost tracking and troubleshooting. For high-throughput scenarios, batching writes and using asynchronous persistence will reduce latency impacts.

For teams building product features, formalize the pattern in a reusable workflow template or custom node: capture the response, persist it to a context store, and expose a simple API for retrieval. Document failure modes, include unit tests for message extraction logic, and monitor conversation drift so the downstream business logic can adapt when the model's output format changes.