A new open-source project called Claude Code OS is bringing persistent operational memory to Anthropic's Claude Code CLI agent, solving one of the most complained-about limitations of AI coding assistants: starting every session from absolute zero. The project dropped on Hacker News this week with a simple pitch—drop it into your project root and never re-explain your stack again.

The Memory Problem Claude Code Never Solved

Claude Code executes tasks well, but it's always been stateless by design. Every new conversation means either dumping your entire project as context (expensive in tokens) or manually re-explaining changes you made last week. Developer bernardohcrocha built Claude Code OS to close that gap with a _brain/ folder that lives inside your project repository and updates itself automatically based on what actually changed. The brain folder contains structured subdirectories: core/ for product knowledge and brand guidelines, operations/ for metrics and customer data, skills/ for permanent rules like fraud-checking logic, and tasks/ for scheduled recurring jobs. On first run, Claude Code OS scans your entire codebase, docs, and configs to build the initial state. From then on, it handles updates itself.

Token-Efficient Daily Syncs via Git Diffs

What makes this approach interesting isn't just persistence—it's how efficiently it stays current. Every day, Claude Code OS runs git diff HEAD~1 to detect exactly which files changed. If you modified three files today out of a thousand-file project, it reads only those three. One file update cascades across every dimension it touches: product knowledge, customer segments, metrics, and skills all refresh simultaneously without a full re-read. This matters for teams running Claude Code on larger codebases where context window costs add up fast. The architecture is MIT-licensed and completely independent from Anthropic—the company makes clear in the repo that Claude Code OS is "an independent open-source project, not affiliated with or endorsed by Anthropic."

Live Dashboard and Scheduled Task Execution

Beyond passive memory, Claude Code OS includes a dashboard.html file that auto-refreshes every five minutes in your browser. The dashboard isn't hardcoded—it builds sections dynamically from your actual data and evolves as your project does. You get proactive suggestions, active project status, scheduled task results, and weekly reviews without opening a chat window. The scheduling system is where this gets genuinely powerful. Users can define tasks like "Every Monday: pull last week's numbers from Stripe, compare against previous week, flag anomalies, and queue follow-ups for accounts below 10% quota." Set it once with natural language instructions. It writes _brain/skills/ files automatically to preserve the rule forever. No config files. No editing JSON. Just tell it what you want done.

One-Command Setup and Cloud Backup

Getting started requires a single curl command: curl -fsSL https://raw.githubusercontent.com/bernardohcrocha/claude-code-os/main/setup.sh | bash run from your project root after launching Claude Code. The setup script scans your entire project, connects to existing tools, and asks only what it can't infer—no forms, no manual configuration files. Optional GitHub CLI integration (gh) automates a private brain repository creation and enables cloud backup with every commit. Wipe your machine, clone the brain repo fresh, and pick up exactly where you left off. The scheduler installs automatically using launchd on macOS or systemd on Linux, catching up on missed tasks after sleep or restart. Requirements are minimal: git and Node.js 18+. Speech-to-text integration works with Wispr Flow or handy.computer for hands-free operation—think "Check if any Pro accounts are near quota" while making coffee.

Key Takeaways

  • Claude Code OS adds persistent operational memory via a _brain/ folder that self-updates daily using git diffs
  • Token-efficient: only changed files trigger updates, not full project re-reads
  • Dashboard auto-refreshes every 5 minutes with evolving sections based on live data
  • Natural language task scheduling: "Every Monday pull Stripe numbers and flag anomalies" becomes permanent automation
  • MIT-licensed open source, no affiliation with Anthropic, zero subscription cost

The Bottom Line

This is the operational memory layer that Claude Code users have been building manually (and badly) for months. The git diff approach is clever—instead of fighting context limits, it works within them by tracking deltas rather than snapshots. Worth watching as it matures, and even more worth forking if you're already deep in Claude Code workflows.