🧠 ChatGPT remembers, Manus sells, Codex develops on iOS

For years, starting a new ChatGPT conversation meant starting from scratch. You had to remind it who you were, what you liked, how you write. OpenAI addressed this partially with a manual "saved memories" feature, but it still depended on you explicitly telling the model what to keep. Most people never bothered.

Today that changes. OpenAI is rolling out a new memory architecture built on top of its "dreaming" system, a background process that automatically synthesizes useful context from your past conversations without you lifting a finger. The model now carries forward your preferences, writing style, and recurring context across sessions on its own.

The numbers back it up. Factual recall has jumped from 67.9% to 82.8%, preference adherence from 55.3% to 71.3%, and accuracy over time from 52.2% to 75.1%, according to OpenAI's internal benchmarks. There's also a new memory summary page where you can review, edit, or correct what ChatGPT thinks it knows about you.

Plus and Pro users in the US get it today. Free and Go users follow in the coming weeks.

Manus, the AI agent platform from Meta, just launched a direct Shopify integration. The pitch is simple: describe the business you want to build, and Manus generates the storefront including branding, navigation, product pages, and checkout routing. Shopify then handles payments, inventory, and fulfillment in the background.

For people who have been putting off launching a store because the setup felt overwhelming, this removes most of the friction. Manus can also read your existing Shopify sales data to spot slow-moving products, build marketing campaigns, write email drafts, and generate social content. All from the same chat window.

The backend stays in Shopify's hands, which is the sensible call. Your product catalog, order management, and customer data stay inside Shopify's dashboard. Manus acts as the creative and operational layer on top.

OpenAI's Codex coding agent just got a plugin that closes a frustrating gap for iOS developers. The new Build iOS Apps plugin lets Codex view and interact with your iOS app directly inside its in-app browser, open SwiftUI previews, and hot-reload edits without you switching between tools.

Before this, even if Codex was writing your Swift code, you still had to manually build and test in Xcode on the side, check what broke, then come back and describe the issue. That back-and-forth added friction to every iteration.

With this plugin, Codex can complete more of the development loop on its own: write the code, preview the UI, catch visual issues, and reload changes, all in one environment. Worth installing if you are working on any SwiftUI project inside Codex.

Honen is an AI-powered course builder designed for the era of workforce reskilling. You drop in your organization's documents, call recordings, handbooks, or onboarding materials, and Honen turns them into structured interactive courses with a built-in AI tutor. The courses improve over time as learners move through them. Honen has partnered with NVIDIA to bring professional-grade AI literacy training to 250,000 learners, and for the first time, NVIDIA's Deep Learning Institute curriculum is available to students under 18. If you manage a team and spend too much time onboarding or upskilling people the manual way, this is worth a look.

Reve 2.0 is a native 4K image generation model with an interesting technical bet underneath it: instead of just interpreting a text prompt and guessing at composition, it builds a structured layout first, defining where objects go, how they relate spatially, and what each element looks like before rendering anything. The result is images that feel designed rather than generated. Reve 2.0 currently sits at number two on the Text-to-Image Arena leaderboard, behind GPT-Image-2 and ahead of Google's Gemini 3.1 Flash Image Preview. If you do creative work that requires precision layout, this is the most controllable image model available right now.

Gemini Omni is Google's new multimodal video model, launched at Google I/O 2026. It takes text, images, audio, or existing video as input and generates or edits short clips through natural conversation. Instead of starting over for each change, you refine the output with follow-up prompts. Here is how to try the self-clone workflow:

Here's the prompt used: "A gigantic version of me wearing the exact same outfit and sharing the same face walks into the scene from the left. We both look at each other and wave enthusiastically."