Browse / Software development / agent-os-starter
agent-os-starter
matthewalberto
A template for a Claude Code workspace where agents do recurring work and learn from corrections: pointer files, one-fact memory, skills and hooks
Open matthewalberto/agent-os-starter → Checked against the four rules.
agent-memoryai-agentsclaude-codellmpythontemplate
Built for Claude Code
ships .claude CLAUDE.md and nothing for other agents.
In its own words
A small template for organising a Claude Code workspace where agents do long-running, recurring work, and get better at it over time instead of repeating the same mistakes.
Why it matters
- The problem: every agent session starts with no memory, so the same corrections get made again and again, and a growing instruction file soon costs more context than it saves.
- What this saves: each correction lands in one durable place (a memory, a skill edit or a hook), and each session loads only what its task needs.
- Where it came from: this is the structure behind my own workspace (about 7,600 directed commits since May 2026), rebuilt here with generic examples.
It's not a framework. It's a folder layout, a few conventions, two scripts and a hook. Copy it, delete what you don't need, and grow it as your work does.
The idea in one paragraph
An agent session starts with no memory of the last one. Anything it learned is lost unless it's written to a file the next session reads. So every recurring job gets a trigger, rules, tools, safety checks and a learning step, and every correction lands somewhere durable: a memory if it's a fact, a skill edit if it's a better way to do the job, a hook if it's something that must never happen. That's the "closed loop". docs/closed-loop.md explains it in full.
Quoting 29% of the README. Read the rest on GitHub ↗
Read from the repository's git tree on 2026-08-17.