Want to build some cool stuff with Gradium AI's voice models? Join our hackathon / builder challenge! We'll drop all the details and specifics in a few days, but reserve your spot now 👇

An AI agent remembering a fact isn't enough. It also needs to understand when that fact changed. Imagine an agent handling customer support. A customer updates their delivery preference three times. A basic memory system might retrieve all three versions without explaining which one is current. Oracle AI Agent Memory 26.8 introduces graph-aware retrieval to solve this problem. ↳ Typed links like supersedes, refines, supports, and contradicts make relationships between memories explicit. ↳ Automatic linking can identify relationships as new memories are extracted, without requiring developers to create every link manually. ↳ The num_hops setting controls how far retrieval follows those relationships, while max_linked_results limits the additional context returned. Older memories aren't simply deleted. They can remain available as historical context, helping agents understand how information evolved. This is useful for long-running agents where customer preferences, support cases, and business facts change over time. The release also introduces image-aware memory and additional enterprise controls. Want to see how it works? The article includes a runnable Python notebook: fandf.co/4e7wkwY — Thanks to @OracleDevs for sponsoring this post.

Hey @grok, what should developers focus on to stay relevant over the next 5–10 years?

We're planning a giveaway, but wanted to ask the community first! What would you like as a prize?

🎃

I just released 90 free, open source skills that teach AI assistants to take notes properly. It's something that's been annoying me for a while. You hand an assistant a lecture, a recipe, or a meeting, and you get back the same old tidy list of bullets every time. These skills fix that. Each one teaches a single job the way an expert would do it. There's one for Cornell notes that actually puts questions in the cue column, which I find most assistants get wrong. The meeting notes one won't turn "we should probably look into that" into a decision no one made. The recipe one keeps the "until the edges go golden" bits instead of flattening everything into "bake 20 min." And there are some for niche stuff too, like birding logs, music practice, and D&D session notes. And none of them makes anything up. If your source didn't say it, it gets marked [?] for you to fill in. Here's why I made them. A year ago I wrote on my blog that the habit of writing things down matters more than the software. Then I wrote every day, hit a few hundred more notes, and ended up building my own notes app because hey I'd been wrong. I build with AI all day, and the faster I go, the more I need to write. It just felt backwards that the AI was writing the worst notes. 90 skills in 4 collections: - note-taking methods - notes for your hobbies - everyday notes - writing They work with Claude, Codex, Cursor, Copilot, Gemini CLI, and more. npx skills add scratchpad-sh/skills Of course they work with any notes app, or none at all, since everything comes out as plain Markdown.

Thank me later. javascriptquiz.com.

Quickly and easily create a CSS grid layout for your website!

I built parley for the @active_builders x @GradiumAI hackathon. Here's the idea. You hold a key, talk to Claude Code, and let go. Your words show up in the prompt already cleaned up, and the prompt gets sent. The "um"s, the "wait, no, Friday" and misheard names are all fixed before Claude sees them. Say "the auth file" and it reaches Claude as @ src/auth.ts. Say just "stop", "undo that" or "clear the chat", and parley presses the keys for you. It talks back too. When Claude finishes its turn with a question, parley chimes and opens the mic by itself, so you just answer. Say "answer me out loud" and Claude's reply gets read to you. Gradium AI does all of the speech. Its speech-to-text streams your words into a little pill at the bottom of the screen while you're still talking. Its text-to-speech reads the answers. You can speak French, German, Spanish or Portuguese, and Claude still gets English. Parley remembers any word you correct so it's spelled right next time. It's one Rust binary: a background process with a mic, Claude Code hooks, and a GTK pill on Linux. It runs on Linux (Hyprland), macOS and Windows. It was a really fun build, and Gradium's speech-to-text is fast enough that talking to Claude feels natural. The repo is at github.com/radiumcoders/parl…

I built heckled.dev for the @active_builders x @GradiumAI hackathon. Here's the idea. You paste a landing page URL, pick a heckler, and get roasted out loud for about 60-90 seconds. Then you get a fix list you can paste straight into your AI to (hopefully) improve your landing page. A headless Chrome screenshots your page and maps the headline, buttons, pricing, and proof. Claude reads all that and writes an eight-part stand-up set, with each bit pointing at the elements it mentions. Gradium AI does the voice and each heckler (Jaded VC, Growth Bro, Creative Director, and Documentarian) is a voice I built with Gradium Voice Design, with its own pronunciation dictionary. It does French roasts too. Gradium also returns word-level timestamps, and those run the whole player. The captions follow the voice, and the camera pans to a spotlight on whatever's being roasted at that moment. It's one Next.js process with a SQLite queue. About a minute from URL to heckle. It was a fun little project, and I got to try Gradium AI's tools, which was pretty awesome. The repo is at github.com/phillips-jon/heck…

typeui.sh is an amazing design layer for AI coding agents! 1. Design skills 2. UI prompts 3. Works with Claude, Codex & Cursor Check it out!

Co-organizing a builder challenge with @joncphillips: Active Builders x Gradium AI. Free, open to everyone, teams welcome. 1st place wins a Raspberry Pi, 1M Gradium credits and Active Builders lifetime membership. Sign up → activebuilders.dev/hackathon…

LFG 🚀 Make sure you grab your spot and join us in this hackathon! Exact details coming in a few days.

Alright, we teamed up with @GradiumAI for a builder challenge / hackathon 🚀 The idea is that you build something real with Gradium. It's free and open to everyone, and you can win some cool prizes. 🥇 First prize is a Raspberry Pi + 1M Gradium credits + lifetime access to @active_builders 🥈🥉 runner-ups get 500k Gradium credits each and lifetime access to @active_builders The brief drops in a few days, but you should save your spot now 👇

We teamed up with @GradiumAI to bring this Hackathon to you. Details below!