forge

A retrospective system for continuous agent improvement analyzes task execution against plans to identify deviations and extract lessons. It operates in retrospective, feed-forward, and mutation modes to store historical data in project memory directories.

Feed-forward advisories consult past lessons to provide targeted guidance before starting new tasks. This mechanism prevents known pitfalls by surfacing historical patterns extracted from previous project retrospectives stored in memory directories.

Generate skill mutation proposals by analyzing accumulated retrospective data to suggest evidence-based improvements to existing skills. This mutation analysis identifies recurring issues and proposes concrete enhancements to improve agent performance over time.

You need project memory directories to run task retrospectives because the system stores historical patterns and lessons learned in them. This local storage enables continuous skill improvement and deviation analysis across multiple task executions.

Run a retrospective analysis after completing significant tasks to identify what went wrong and what was learned. This ensures continuous improvement by capturing deviations and extracting actionable lessons while the task execution context is available.

You can use retrospective lessons to prevent known pitfalls in planning sessions through feed-forward advisories. The system surfaces past lessons before new tasks, advising the agent to avoid previously identified rabbit hole scenarios and recurring issues.