Ratchet is a lightweight framework for building AI agents that continuously improve themselves — autonomously, over time, with minimal human involvement.
Agents persist and curate knowledge across sessions. Long-term memory stays lean and current. Daily logs feed weekly synthesis.
Every failure gets a postmortem. Root cause, blast radius, prevention tasks. "Fixed" means the class of problem is addressed — not just the instance.
Self-identified tasks execute during heartbeats. P1s run immediately. The agent works through the backlog without prompting.
Authoritative state for location, timezone, and preferences. Never assumed — always checked before creating anything time-sensitive.
Trust expands over time as it's earned. The framework is explicit about what the agent decides alone vs. what it escalates.
Ratchet is conventions, not code. Wire it to OpenClaw, Claude Code, Cursor, or any agent platform. Write an adapter, contribute it back.
your-agent/ ├── MEMORY.md # Long-term memory — curated, evolving ├── BACKLOG.md # Self-directed work queue — P1/P2/P3 ├── context.json # Authoritative state (location, units, prefs) ├── incidents/ │ ├── README.md # Postmortem format spec │ └── INC-001-*.md # Incident logs ├── memory/ │ └── YYYY-MM-DD.md # Daily logs └── bin/ # Utility scripts