| name | self-improvement |
| description | Captures learnings, errors, and corrections to enable continuous improvement. Use when a command fails unexpectedly, the user corrects a misunderstanding, an external tool breaks, or a better approach is discovered. 适用于“把这个报错记下来”、“记住我刚才说的规则”、“更新最佳实践”或“把这个教训沉淀下来”等需要自我进化的场景。 |
Self-Improvement
Capture development insights, command failures, and user feedback into Trae's Core Memory, MCP Memory (for cross-session persistence), or project-level rules.
Classification rule of thumb: Core Memory for session-level corrections; MCP Memory for cross-session reusable knowledge; Project Rule for recurring patterns worth codifying; New Skill for a capability that deserves its own toolbox.
Use This Skill
- Command/Tool Failure: An operation fails unexpectedly (e.g., port conflict, missing dependency).
- User Correction: The user corrects your output, logic, or architectural approach.
- Best Practice Discovery: You find a more efficient or robust way to perform a recurring task.
- Knowledge Gap: You realize your understanding of the project or an API is outdated.
- Task Wrap-up: Before completing a major task, review if any persistent workarounds should be captured.
Do Not Use
- Transient network glitch or environment fluke that won't recur.
- Non-technical chat without actionable feedback.
- Routine code change without any new learning involved.
- Task progress summarization (use standard output contracts for that).
Quick Reference
| Situation | Action |
|---|
| New learning discovered | resources/execution-guide.md → classify → store |
| Recurring pattern (3+ times) | resources/promotion-guide.md → consider Rule |
| Learning qualifies as reusable Skill | resources/execution-guide.md → use skill-creator |
| Memory is near limit or outdated | resources/memory-maintenance.md → audit → DELETE |
| User says "帮我整理记忆" | Ask scope → explain snapshot limit → user verifies total from settings → resources/memory-maintenance.md 用户触发流程 |
| Memory fragments seem to be growing | Ask user → if confirmed, use resources/memory-maintenance.md 整合审计 |
| Similar memory already exists | Use UPDATE on existing entry, don't duplicate |
Output Format
When you finish logging a learning, report back using this format:
**教训已记录 (Learning Captured)**
**分类:** `[Core Memory / MCP Memory / Project Rule / New Skill]`
**摘要:** [一句话概括]
**应用场景:** [什么情况下会用]
After reporting, resume the original workflow that triggered the learning. Do not treat self-improvement as a terminal step — the caller's execution should continue after logging.
Failure Strategy
- Unclear Correction: Ask for clarification before logging.
- Redundant Memory: Use UPDATE to merge with existing; don't create duplicates.
- Sensitive Data: NEVER log secrets, tokens, or private keys.
- Low-Value Learning: Skip trivial/obvious insights to avoid memory pollution.
- Tool Unavailable: If
manage_core_memory fails, fall back to manual capture via resources/execution-guide.md; if any script fails, retry once before reporting failure.
Integration
systematic-debugging: Client — non-obvious root causes are logged via self-improvement as knowledge_gap or insight.
test-driven-development: Client — false-green tests discovered during mutation testing are logged as anti-pattern experiences.
verification-before-completion: Client — non-obvious validation failures are logged as Experience.
finishing-a-development-branch: Client — during branch wrap-up, persistent workarounds and rule discoveries are promoted via self-improvement.
subagent-driven-development: Client — repeated subagent failures are logged to prevent future mis-routing.
memory-kernel: Downstream — after capturing a learning, if classified as MCP Memory, invoke memory-kernel to write it to the knowledge graph for cross-session persistence.