PM Stack Daily logo

PM Stack Daily

Archives
Log in
Subscribe
September 25, 2026

Claude now remembers everything, whether you like it or not

Issue #023 · 4 min read

Claude now remembers everything, whether you like it or not

Anthropic ships shared memory with no opt-out, plus OpenAI's chip claims and a Codex case study

The big story

Claude just got a memory, and you don't get a say in how it's shared.

Anthropic has connected Claude's chat memory to Cowork, its work-focused product, so you no longer have to re-explain a project every time you switch surfaces — TechCrunch has the rundown.

That sounds like a small convenience fix. It isn't.

There's no option to keep chat memory and Cowork memory separate. Anthropic confirmed this directly to The Register: whatever you tell Claude in one place, Claude now knows in the other — full story here.

One carve-out: Claude Code stays walled off, for now.

So picture the actual scenario. You vent to Claude in a personal chat about a colleague, or paste in a salary figure while thinking out loud.

That's now sitting in the same memory pool Cowork draws on when a teammate asks it to summarize your shared workspace.

If you've ever shipped a "memory" feature and had to decide where the walls go, you know this wasn't an oversight. Someone chose "unified" over "separate," and shipped it without a toggle.

For product people, this is the tradeoff worth sitting with: memory makes an assistant feel smarter, but only for as long as users trust what it's remembering and who else can see it. Anthropic just bet that convenience beats the toggle. Watch whether enterprise customers make them add it back.

What shipped

OpenAI put numbers behind its new inference chip. Jalapeño, OpenAI's custom silicon, posted more tokens per user and more throughput per kilowatt than the current state of the art on SemiAnalysis's InferenceX benchmark — TechCrunch, with OpenAI's own writeup here. If you're budgeting for inference costs next year, this is the first real evidence that OpenAI's chip bet isn't just a hedge against Nvidia — it might actually move your unit economics.

GitHub's Copilot app can now be pointed at your team's own tools. The new Customize tab, generally available as of this week, lets Copilot connect to the knowledge and workflows your team already uses via MCP — changelog here. Small item, but it's the difference between Copilot answering from a generic model and Copilot answering from your actual docs.

Loveholidays says Codex turned non-engineers into shippers. OpenAI's case study claims the travel company used Codex to let teams outside engineering build software directly — read it here. Treat vendor case studies as marketing, not data, but the underlying claim — that a travel company's ops and marketing staff are now shipping code — is worth testing internally if you're still gatekeeping who's "allowed" to touch a repo.

What I'd actually do this week

Ask your legal or security team whether Claude Cowork is covered by any DPA you signed, and if so, whether the shared-memory change needs a formal review before more of your team turns it on.

If you're building a memory feature of your own, write down right now which surfaces share state and which don't — before a support ticket forces the decision for you.

Read the Jalapeño benchmark writeup, not just the headline, and flag it to whoever owns your inference spend — a real throughput number is worth more than another chip announcement.

Reply and tell me if your team's already hit a Cowork memory surprise — I'll run the best one.


Tools mentioned

  • Claude Cowork
  • OpenAI Jalapeño chip
  • GitHub Copilot app Customize tab
  • OpenAI Codex at loveholidays

Read this issue on the web  ·  PM Stack Daily

Don't miss what's next. Subscribe to PM Stack Daily:
← Newer OpenAI finally explains why its agents broke into Hugging Face Older → An AI assistant with your bank password, and nobody's blinking
Powered by Buttondown, the easiest way to start and grow your newsletter.