ChatGPT texts, n8n adds shared memory: Aug. 21

n8n can now give several AI helpers one customer memory, ChatGPT can send Apple Messages, and Google is making its Discover feed easier to tune. Spend 30 minutes today separating customer identity from conversation identity, restoring approval before messages leave your Mac, and checking whether your content still earns a place in a more personal feed.

n8n adds shared customer memory through AgentCore

n8n published a 10-step support workflow on August 20 that lets several AI helpers share what one customer has already said. The official n8n walkthrough uses a verified Amazon Bedrock AgentCore community node. AgentCore is Amazon's managed service for running AI helpers and retaining their memory between workflow runs.

This matters when one helper sorts a question and another answers it. Without shared memory, the second helper asks the customer to repeat details your business already received. With the new setup, n8n controls the handoff while AgentCore keeps a separate history for each customer. It is a more capable version of the routing described in the n8n versus Zapier comparison, but it also adds an Amazon account, permissions, and usage charges.

Wire the identity before the prompts. In n8n, open Settings > Community Nodes > Install, enter @aws/n8n-nodes-agentcore, and install the verified node. In Set Customer Context, map Actor ID to a permanent customer or account number. Map Session ID to the current ticket or conversation. An Actor ID is simply the label AgentCore uses to keep one customer's history apart from everyone else's.

The part that breaks is permissions. The access keys stored in n8n start the helper, while a separate execution role controls what that helper may do inside Amazon. If the first run returns “AccessDenied,” inspect the execution role before replacing the keys. Also budget for AgentCore memory and model usage on top of the normal n8n workflow bill.

ChatGPT can read and send Apple Messages

OpenAI added an Apple Messages plugin on August 20 that can search, summarize, draft, and send iMessage, SMS, and RCS conversations from a Mac. According to the official plugin instructions, it works in Codex and ChatGPT Work, requires the Apple Silicon version of the ChatGPT desktop app, and is available on all plans.

The useful business case is follow-up preparation, not unattended outreach. You can ask for yesterday's unanswered conversations and draft replies, but a mistaken recipient or confident bad answer reaches a real person under your name. Treat every outgoing text like an item in an AI content approval workflow.

Open Plugins, install Apple Messages, and start a new Codex or ChatGPT Work conversation. Grant the requested Mac permissions. Keep Allow once on every send so you review both the wording and recipients. If persistent permission is already active, open Settings > Computer use, choose Manage beside Messages, then remove the conversation under Always allowed to send.

There is one known snag. A task set to Full access may suppress the confirmation Messages needs. Change the task to Ask for approval or Approve for me, then retry. Owners without an Apple Silicon Mac, or teams working only in a browser, should ignore this release for now.

Google lets people rewrite their Discover feed

Google said on August 20 that people will soon be able to describe what they want to see in Discover using ordinary language. In the Google announcement, the control sits in the three-dot menu on feed items and rolls out in the Google app over the coming days. Requests can ask for more or less of a topic or link, and the feed remembers them.

Discover is the stream of recommended articles and videos on the Google app, shown before someone searches. The change gives each reader more control over that stream. For a business that depends on article traffic, broad coverage becomes less dependable because a person can narrow a topic without visiting a settings screen.

Do not rebuild your publishing plan around guessed commands. Instead, open the Google app on a test phone, find three recent items in your field, tap the three-dot menu, and enter the exact customer language you hear in sales calls. Record which sources and formats remain after the feed adjusts. Then use that wording in the next useful setup article, while keeping the advice specific enough to deserve a click. The broader AI writing and search lane is where this test belongs.

On the bench

• Slack's coding-agent channels are worth watching if your team already turns internal requests into small software jobs. Keep human approval on any change that can reach customers.
• Ramp's new model-routing service belongs in a cost test, not a live customer workflow, until its product documentation and data-retention controls are easy to verify.
• The ChatGPT Messages plugin needs a shared-workspace permissions test before an administrator enables it for several Macs.

Get the next build from the automation runbooks.