No More AI Amnesia: How Google's Own Antigravity Setup Finally Made My Assistant Remember Everything

Let's be honest, we've all been there. You're deep into a project with your AI assistant, making real progress, and then you step away for a bit. You come back, fire it up again, and it's like meeting a stranger. All that context, all those specific details you meticulously fed it? Gone. Poof. It's incredibly frustrating, a productivity killer, and frankly, it makes you wonder if these powerful tools are really as smart as they claim.

That was my daily struggle with Google Antigravity, a tool I knew had immense potential, but kept tripping over its own memory gaps. I’m Nolen Jonker, and after years writing about tech and creativity, I’ve seen my share of AI quirks. This particular one, the constant forgetting, was starting to wear me down. It felt like I was constantly onboarding a new assistant for every single session. But then, I stumbled upon Google's own recommendations for Antigravity's setup, and it was like a lightbulb moment. What I found wasn't some arcane hack, but a foundational approach using 'Rules' and 'Skills' that completely transformed my interaction with the AI.

Antigravity, for those unfamiliar, is a remarkably capable AI tool. But its true power, I discovered, lies not just in its raw processing, but in how you configure it. Google provides a 'Customizations' panel within Antigravity, offering six distinct surfaces: MCP, Skills, Rules, Plugins, Hooks, and Sidecars. For me, the magic truly began when I dove into 'Rules' and 'Skills'. These weren't just obscure settings; they were the key to teaching Antigravity about my world, in a persistent way.

First up: Rules. Google's documentation describes these as "manually defined constraints for the Agent to follow." Sounds a bit dry, right? But think of it this way: these are the fundamental truths, the ground rules you lay down for your AI. They're actually implemented as plain markdown files, which Antigravity automatically injects into every session. It's brilliant in its simplicity. You can have global rules, like a `GEMINI.md` file in your `~/.gemini` directory that applies to everything, or you can get more specific with workspace-level rules stored in a `.agents/rules` folder at your project's root.

The beauty of workspace rules is their flexibility. You can set them to be 'Always On,' triggered by relevance (meaning Antigravity intelligently decides when they're applicable), manually activated with an `@mention`, or even tied to specific glob patterns. For my setup, especially concerning my Obsidian vault, this was a game-changer. I used Rules to define my entire environment: the tools I use, my specific folder conventions, the naming patterns I prefer for files, the actions I generally want the agent to take, and crucially, even areas it should absolutely not touch. Suddenly, Antigravity wasn't just a generic AI; it was my AI, understanding the nuances of my workflow. It was a genuine relief to see it consistently apply these guidelines without me having to type them out every time.

Then there are 'Skills.' These, Google explains, are "reusable packages of knowledge" that operate on a progressive disclosure pattern. This is genius for preventing 'context rot' and keeping things snappy. The agent doesn't load all the instructions for every skill upfront. Instead, it only sees the skill's name and a brief description. The full, detailed instructions only load when your prompt semantically matches that description. It's like having a library of specialized manuals, but the AI only pulls out the relevant one when you actually ask for something related to it, rather than trying to digest the whole library at once. This approach, outlined in Google's own codelab, is key to maintaining both efficiency and relevance.

Setting up a skill is straightforward enough. You store workspace skills in a structured path: `.agents/skills/skill-name/SKILL.md`. Inside that `SKILL.md` file, you start with some YAML frontmatter defining the skill's name and, most importantly, its description. This description isn't just for you; it's the trigger for the AI. It needs to be written in the third person and include phrases or keywords that are likely to activate the skill. I crafted a skill specifically for recurring chores within my Obsidian vault. Now, instead of painstakingly outlining the steps each time, I just refer to my custom skill, and Antigravity knows exactly what to do. It’s efficient, it’s intelligent, and it feels truly collaborative.

The difference these configurations have made is monumental. My Antigravity assistant no longer suffers from frustrating short-term memory loss. It remembers my preferences, my project structure, and my unique ways of working, session after session. This isn't just about convenience; it's about a fundamental shift in how I interact with AI. It’s moved from a constant struggle of re-education to a fluid, productive partnership. If you're using Antigravity and battling with a forgetful AI, take it from me: diving into Google's own recommendations for 'Rules' and 'Skills' is the smartest move you can make. It truly stops the forgetting, allowing you to focus on what really matters.