A contrastive entry in corrections.md. What went wrong, then the rule that prevents it from happening again.
A correction is how Arkeus learns from failure. It is a contrastive entry in corrections.md with a specific shape: a dated incident describing what happened, followed by a rule stated in a form that the system can apply later.
The contrastive shape is important. A correction is not a lesson in the abstract. It is an incident paired with the rule derived from that incident, so that when the next similar situation arises, the rule is grounded in a concrete memory the system can pattern-match against. The format enforces specificity: no generic advice, no vague improvements, no optimism.
Corrections come from two paths. The reactive path is mid-session: Ryan corrects an agent in real time, and the correction lands in corrections.md as a new entry. The proactive path is end-of-session: patterns that showed up during the session get pulled out and written as contrastive rules that apply to future sessions.
The ceiling on corrections.md is around 4,000 tokens. When the file exceeds that, older entries that have been superseded by newer ones get pruned. Agent-local corrections stay in each agent's own file and only graduate to the global corrections.md when the same mistake appears across three or more agents.
Corrections are the feedback loop that makes the system anti-fragile. Every time the system fails, it writes down what failed and the rule that would have prevented it. The next cycle loads that rule at the top of context. The failure does not repeat. This is what separates Arkeus from systems that apologize and continue doing the same thing.