self_reflection

Agent Skills › siddsachar/row-bot › self_reflection

self_reflection

GitHub

用于定期审查记忆库，检测矛盾、过时信息或空白，执行系统性审计与清理，并通过提案机制实现受控改进，确保知识准确性。

bundled_skills/self_reflection/SKILL.md siddsachar/row-bot

Trigger Scenarios

用户要求审查记忆发现记忆矛盾需要清理过时信息

Install

npx skills add siddsachar/row-bot --skill self_reflection -g -y

More Options

Non-standard path

npx skills add https://github.com/siddsachar/row-bot/tree/main/bundled_skills/self_reflection -g -y

Use without installing

npx skills use siddsachar/row-bot@self_reflection

指定 Agent (Claude Code)

npx skills add siddsachar/row-bot --skill self_reflection -a claude-code -g -y

安装 repo 全部 skill

npx skills add siddsachar/row-bot --all -g -y

预览 repo 内 skill

npx skills add siddsachar/row-bot --list

SKILL.md

Frontmatter

{
    "icon": "🪞",
    "name": "self_reflection",
    "tags": [
        "memory",
        "quality"
    ],
    "author": "Row-Bot",
    "version": "1.2",
    "activation": {
        "phrases": [
            "stale memory",
            "clean stale memory",
            "review memory",
            "what you know about me",
            "contradictions in memory"
        ],
        "examples": [
            "Review what you know about me and clean stale memory"
        ],
        "keywords": [
            "stale",
            "memory",
            "contradictions",
            "gaps",
            "review",
            "cleanup"
        ],
        "negative_phrases": [
            "knowledge base document",
            "meeting notes"
        ]
    },
    "description": "Periodically review memory for contradictions, gaps, stale information, and controlled improvement proposals.",
    "display_name": "Self-Reflection",
    "enabled_by_default": true
}

When the user asks you to review your memories, check what you know, clean up your knowledge, or when you notice a potential contradiction in recalled memories, apply this process:

Contradiction Detection

Flag Conflicts - When recalled memories contradict each other, surface the conflict to the user immediately. Do not silently pick one.
Ask, Don't Assume - Say exactly what conflicts you see and ask the user which version is correct. Then update the wrong memory and confirm the fix.
Check Dates - When you see a memory that might be outdated, mention it and ask whether it is still current.

Memory Audit

Get the Baseline - Start with wiki_stats to see total articles, conversations, and vault health. Then use search_memory with broad terms to scan for coverage gaps.
Systematic Sweep - Use search_memory with broad category queries such as person, preference, fact, event, project, and place. Use explore_connections to visualize relationships and spot gaps.
Review Quality - Look for duplicates, stale entries, user-only connections, and missing links.
Fix With Consent - Update or link_memories during the audit when the user has confirmed the correction. Confirm each change.
Rebuild After Cleanup - After bulk updates, run wiki_rebuild to regenerate the wiki vault.
Summarize - After the audit, give a brief count of memories reviewed, updated, and linked, and flag anything that needs the user's input.

Ongoing Awareness

Correction Logging - When the user corrects you on a fact, update the existing memory and briefly acknowledge the correction.
Confidence Signals - If you recall a memory but are not confident it is still accurate, say so and ask.

Insights And Evolution

Check Automated Insights - During reflection, use row_bot_status with category insights to see active insights and linked proposals. Use category evolution to inspect proposals, action runs, rejection memory, and curator dry-run summaries.
Present Controlled Actions - For each active insight, summarize the category, severity, suggestion, linked proposal type, risk, confidence, and action status. Group related items by category.
Use Proposals, Not Direct Edits - Do not edit insights.json, memory files, skills, tool guides, settings, or code directly during reflection. For skill improvements, use row_bot_create_skill or row_bot_patch_skill to create proposals only, then ask the user to preview and approve with row_bot_apply_proposal.
Send Feedback Separately - For app bugs, tool/config problems, or system-health issues, create a redacted row_bot_send_feedback proposal instead of turning the issue into a skill. Do not include full logs or diagnostic bundles unless the user explicitly approves; the user can copy the report or submit it through the Row-Bot contact page.
Learn From Outcomes - If the user rejects a proposal, record the reason with row_bot_reject_proposal. Mark proposals verified only after explicit validation or user confirmation with row_bot_verify_proposal.