Agent Experience Refinement

What It Solves

Agents often re-learn the same mistakes across similar tasks: they may notice a failure once, but the lesson is not stored in a searchable, comparable, reusable format. This skill is aimed at Agent post-task review and knowledge capture, turning important decisions, failure causes, and effective practices into structured cards that can be retrieved in later work.

How It Works

The core flow covers experience extraction, card generation, relation-based recommendation, and search-based reuse.

• Input sources: manual task description, conversation history, MEMORY.md / TOOLS.md, or imported .txt / .md files.
• Experience cards: structure task context, trigger conditions, key steps, pitfalls, and value level for later comparison and retrieval.
• Relation recommendation: finds related experiences based on tag overlap, task-type similarity, skill overlap, keyword matching, and scenario similarity; recommendations are usually capped at five to avoid overload.
• Search and deduplication: retrieval uses minimum similarity and result-count constraints; new entries can be checked for high similarity, with merge, replace, or keep options.
• Export and management: supports exporting the knowledge base to Markdown, JSON, CSV, and PDF, with filtering by tag, time, or value level.

Boundaries and Caveats

This skill fits best when the Agent already has multiple task logs, conversation histories, or note files. If the experience base has fewer than three entries, automatic recommendations are limited. It should not interrupt users who are asking rapid follow-up questions or only want the current result. Experience cards are methodological summaries, not complete execution records, so they should still be judged against the current task context.