OpenClaw Experience Distiller

What it solves

During long AI sessions, agents often hit errors, roll back, and retry several times before solving the task on their own. Naive logs mix failure noise with the final successful path, making it hard to reuse the behavior that actually worked later. This skill extracts the “how it was resolved” signal from noisy session history and stores it as searchable experience entries.

How it works

The skill is built around OpenClaw session compaction and hooks. When a long session triggers compaction, the resulting summary removes redundant detail and becomes a better source for experience extraction. The skill then looks for error keywords and final success states to identify the “self-healed error” pattern.

• Manual extraction: when asked to summarize successful experience, it reviews the session or reads the compaction summary, writes a Markdown record under success/, and updates success/index.xlsx.
• Automatic extraction: hooks on session:compact:after and command:new / command:reset can check whether the session ended in a self-healed pattern and persist the experience automatically.
• Lookup and reuse: before complex tasks, it reads success/index.xlsx, matches by keyword or scenario, reuses a known path if found, and records a new one if not.
• Important records: meetings, decisions, and configuration notes can be archived to import/important-records.xlsx with fields such as type, title, summary, date, and status.

Boundaries

This skill fits agent workflows with ongoing sessions, error recovery, and recurring tasks. It depends on OpenClaw compaction settings and hook installation, and it assumes the model can read and write success/, import/, and the index files. Because it keeps only successful paths, it is not a failure-case library, and extraction quality can be limited if the compaction summary omits critical context.