This week's edition covers Anthropic's new memory and Agent Skills APIs for building agents, Karpathy's transparent LLM training pipeline, on-device inference with Windows ML, and circuit-based interpretability tools that cut data requirements by 150x.
Imagine you're building something with Claude, and you want it to remember things from one conversation to the next.
Every session is a first date—charming at first, exhausting by the tenth time you explain your job
Anthropic has just released a memory tool for their Claude API that lets you do exactly that. Instead of Claude forgetting everything between sessions, you can now give it a way to store and recall information, using files you manage on your own system.
So what's new? Now, Claude can write things down, look them up, change them, or throw them away, all in a special memory folder you set up. This memory lives on your own infrastructure, not Anthropic's, and you're the one in charge of how it works.
The 'not our problem' approach to data management, elegantly packaged as user empowerment
The feature rolled out in beta for Claude Sonnet 4, Sonnet 4.5, Opus 4, and Opus 4.1. If you want to use it, you'll need to teach your code how to handle memory—how to save, change, or erase files. In Anthropic's own tests, giving Claude this kind of memory (along with context editing) made it 39% better at certain tasks.
We've quantified the improvement without specifying what "certain tasks" means—a time-honored tradition in product launches
Not bad for a new trick.
Who actually needs this? If you're building an agent that has to remember things over many sessions—like a support bot that keeps track of conversations for weeks, or a coding assistant that remembers why you made certain choices—this is for you. It's for anyone whose projects are too big to fit in a single memory window.
The digital equivalent of walking into a room and forgetting why you're there, except the room costs money per square foot
But if you're just making one-off calls, or your users always give all the context up front, you can skip this. The real question is: Are you tired of sending the same background info over and over, burning through your token budget just to remind Claude what's going on?
We've reached the point where remembering things has become a line item in the budget
Of course, there's a catch. You'll need to roll up your sleeves and build the storage yourself—whether that's a file system, a database, or something in the cloud.
Translation: this feature works great, assuming you do all the hard parts
You'll also have to decide who gets access to what, keep up with changes in the beta, and make sure your data is safe.
Because all the memory lives on your side, Anthropic never sees it. You're in charge of how long things are kept, how they're locked down, and who can peek inside.
Plausible deniability as a feature—they can't be responsible for what they never see
Anthropic gives you some starter tools for Python and TypeScript, but you're the one building the foundation.
Why does this matter? If you're building agents that need to remember things for the long haul, this is how you keep their memory sharp without stuffing everything into context. Context editing lets you sweep away the clutter, but memory lets you hold onto what's important.
With both memory and context editing, agents can finally finish workflows that used to fall apart when they ran out of space—and they do it while using far fewer tokens. If your agents are running for days or making dozens of tool calls, this is a game changer. If not, you can probably wait until you need it.
More AI everywhere doesn’t mean better results. Small teams of AI experts beat mass licenses. Attackers now use AI to run cybercrime at machine speed. Leaders can’t keep up. Tools still forget us. The winners? Focused expertise and hybrid AI, not AI everywhere.
You've built something and you need to know if it works. So you do what's sensible—you ask an LLM to grade it. Factual accuracy, code quality, agent outputs. The machine judges the machine, and you get a number you can act on. Except that number
Imagine you have a bunch of teams, some with AI, some without, and some where everyone gets their own AI. Researchers ran a big experiment with over 400 people to see what actually happens when you mix and match humans and AI in different ways.
Here’s what they found:
