ChatGPT's silent app swap, agents that remember, and free AI video in Google Vids

🔥The one thing to know

🛠Tools and releases

• Gemini Omni 1.2 Flash in Google Vids — Google's video model now runs inside Google Vids, and Workspace users get three 15-second AI videos a month at no cost. Feed it a model photo plus product shots and it holds faces, garments and packaging type consistent across camera angles — branded motion clips become a lunch-break task instead of a vendor job. On the API, the price dropped 30% to $0.07 per second for 720p video, and you can now refine a clip over several chat turns instead of regenerating it from scratch. Google blog

🎨Designing for agents

• The components of agent memory — Anthropic published a walkthrough of how it gives Claude lasting memory: notes the agent writes as it works, its own judgment of what is worth keeping, and "dreaming" — idle-time sessions where the agent re-reads its logs and corrects its own notes. If you design an agent product, memory is now a UX surface: what the agent saves, forgets and mis-remembers is behavior your users experience and will blame your design for. In Pokémon runs, an older model wrote 413 memory files and won one badge; a newer model kept nine files and won four — hoarding context costs more than curating it. Anthropic also found that fixing wiki structure so terms were quoted word for word lifted graded brief performance from 72% to 100%. Anthropic

  “We've found that it helps to tell Claude what memory is for, but not how to use it.”

• Muse Active Memory — Meta is rolling out Active Memory to Muse and Meta AI: the assistant now tells temporary information from permanent information, and recalls earlier context when it becomes relevant again. This is a big-audience, shipped example of memory built as a designed capability — the pattern to study before your product's agent starts remembering things on its own. Meta reports a 2.4x reduction in forgetting critical information versus the previous models. Separately, Muse "spaces" — app-like screens the assistant builds from a description, such as a shopping hub or weekly planner — are rolling out to all Muse users. Meta AI Research

🌍The big AI stories today

• Claude usage credits — Anthropic told Claude customers that usage credits bought on or after December 7 expire twelve months later, with unused amounts removed — a reason not to stockpile prepaid credits for agent work. Hacker News

• Nvidia — A US intelligence memo, first reported by the New York Times, says China's Ministry of State Security recruited a Singapore-based tech executive to spy on Nvidia CEO Jensen Huang — the chip geopolitics behind the tools you run on. TechCrunch

• Public First — The Anthropic-backed AI-safety super PAC Public First spent over $63 million in the midterms and ten of its seventeen congressional candidates won — AI regulation now has a real campaign wing. Techmeme

💡One takeaway