This started with AntiGoldfishMode, a CLI to solve the "goldfish memory" problem in AI coding assistants. That success led me to a bigger question: what if we built a complete operating system designed for AI agents?

AgentOS is my early exploration of this concept. It's currently a TypeScript-based system with three core engines:

Memory Engine: Persistent SQLite-based storage that lets AI agents remember context across sessions. No more losing project knowledge between conversations.

Execution Engine: Sandboxed code execution environment where agents can safely test their suggestions before recommending them. Currently supports multiple languages with Docker integration planned.

Learning Engine: Basic pattern recognition that tracks what works and what doesn't, building simple predictive models from execution outcomes.

Current State: This is very much an early-stage exploration. The core functionality works - agents can store memories, execute code safely, and learn from results. The CLI is functional with commands like agentos memory store, agentos execute, and basic learning capabilities.

The Philosophy: Instead of throwing larger context windows at AI agents (the "context window arms race"), I'm exploring purpose-built tools that give them the specific capabilities they need: memory, execution verification, and learning from experience.

What's Next: I'm working toward an alpha release and would love feedback from the HN community. Are there specific capabilities you think AI agents need most? What would make them better coding partners?

The code is available on GitHub: [https://github.com/jahboukie/agentos.git]

I'm here to discuss the approach, technical implementation, or what you think the biggest needs are for truly AI-native development tools.