Anthropic has merged Claude's memory into one shared memory across Claude chat and Claude Cowork. Until now the two kept separate notes, so anything you'd built up in chat had to be re-explained the moment you handed a job to Cowork. That's gone.

Two other changes matter more than the merge itself:

To Anthropic's credit, the whole thing is visible. Open Settings > Memory and you'll see what Claude has kept as a list of short topic files you can read, edit or delete one at a time. Health, religion, politics and ethnicity are excluded by default unless you flip a toggle, and things like government ID numbers are never stored at all.

Why this matters if you run a business: memory is where your client names, pricing, staff issues and half-formed plans end up. A Claude that remembers all of that across every tool is a real time saver. It's also a lot of context sitting in one place, gathered without you pressing a button.

My honest read: keep it on, but audit it. Most people will never open that settings page, and "on by default" is doing a lot of work in this announcement.

Try this: paste this into Claude today:

Then prune the list in Settings > Memory. Ten minutes, once, and you know exactly what it's carrying around.

One more from the same week: Cowork on desktop now has its own built-in browser (Pro, Max and Team, rolling out from 26 August). It runs separately from your personal browser, so Claude never sees your tabs, bookmarks or saved passwords. You can import logins site by site if you want it working inside a portal, and banking, email and single sign-on sites are excluded by default. Sensible defaults. Keep them.

Since 25 August, ChatGPT Work (OpenAI's agent mode, on paid plans, web and mobile) can sign into websites while it works. When its cloud browser hits a login page it pauses, shows you a secure sign-in form, and you type your details and do any two-factor step yourself. OpenAI says the model never sees your username or password and doesn't store them.

The examples OpenAI gave are consumer ones: booking appointments, checking a reimbursement, setting up utilities. The small business version is more interesting: supplier portals, insurance renewals, government forms, the accounting add-on nobody has an integration for.

The fine print: once you're signed in, that session stays live for future tasks until it expires. The password is protected. The logged-in session it unlocks is the thing doing the work, and it lives on OpenAI's side until you clear it under Settings > Cloud browser > Browser data. ChatGPT does ask before anything hard to reverse, like a booking or a payment.

The rule I'd use: only hand it logins you'd give a new casual staff member on their first day. Supplier portal, yes. Banking, payroll, your main email, no. And clear the browser data after any one-off job.

Read OpenAI's cloud browser guide →

On 27 August, OpenAI published an open letter called A Call for Collective Action on Cyber Defense, co-signed by more than 100 organisations including Anthropic, Google, Microsoft, AWS, CrowdStrike, Cloudflare and several big banks. The core warning: AI-enabled attacks will become "far more widespread and sophisticated" in the coming months, and the usual defences aren't enough.

It landed the same week Taiwanese security firm TeamT5 reported that Chinese state-linked hacking groups more than doubled their attack volume after adding AI to their toolkit, using DeepSeek to write break-in code, attack email systems and map out targets.

You are not a hospital or a water plant. But you are cheaper to attack, and AI has just removed the attacker's biggest cost: personalising the con. The tell we all relied on, the badly written scam email, is gone.

What this could mean for you this week:

Not glamorous, but the letter's own advice to big companies boils down to the same thing: fix the obvious weaknesses first.

Read the open letter →

Google added travel tools to AI Mode in Search on 27 August. Ask it to "track these flight prices for me" and it will email you when fares move, pulling from 300+ airlines and travel sites and working in more than 180 countries, Australia included. It can also show what a flight or hotel costs in points for programs like Hilton, American Airlines and Wyndham, with more being added.

Hotel booking inside AI Mode is US-only for now (English, paid through Google Pay), so that part doesn't reach us yet.

Google Flights has tracked prices for years. What's new is doing it in plain language, mid-conversation, while you're already asking about the trip. Small thing. Also the kind of small thing that quietly replaces a browser tab you've had open for a week.

Try this: open AI Mode in Google Search and paste your next work trip:

Read Google's announcement →

Every story this week is about giving an AI a little more access. On last Friday’s live call we took that one step further: from one AI assistant to a team of specialist agents, with Grok Bot as the worked example.

We covered giving each bot a defined role instead of rewriting the same prompt every week, organising agents into departments with one lead bot coordinating the work, teaching a bot a process by demonstrating it once, and setting approval boundaries before anything is allowed to act on its own. We also compared where Grok Bot, Claude Cowork, ChatGPT Work and a few others each fit.

The discussion that stuck with me was about rolling AI out across a team: saving someone an hour shouldn't automatically mean handing them another hour of work. Decide where that time goes.

I would recommend: don't build twelve bots on day one. Start with one or two clear roles, connect only the tools they need, run a supervised test, and write down what needs your approval. Add the next agent only when there's a clear reason for it.

The full recording and the bot directory we used are in the Lab.