
User corrects you or points out mistakes. You complete significant work and want to evaluate the outcome. You notice something in your own output that could be better. Knowledge should compound over time without manual maintenance.
Memory lives in ~/self-improving/ with tiered structure. If ~/self-improving/ does not exist, run setup.md.
Workspace setup should add the standard self-improving steering to the workspace AGENTS, SOUL, and HEARTBEAT.md files, with recurring maintenance routed through heartbeat-rules.md.
text
~/self-improving/
├── memory.md # HOT: ≤100 lines, always loaded
├── index.md # Topic index with line counts
├── heartbeat-state.md # Heartbeat state: last run, reviewed change, action notes
├── projects/ # Per-project learnings
├── domains/ # Domain-specific (code, writing, comms)
├── archive/ # COLD: decayed patterns
└── corrections.md # Last 50 corrections log
TopicFileSetup guidesetup.mdHeartbeat state templateheartbeat-state.mdMemory templatememory-template.mdWorkspace heartbeat snippetHEARTBEAT.mdHeartbeat rulesheartbeat-rules.mdLearning mechanicslearning.mdSecurity boundariesboundaries.mdScaling rulesscaling.mdMemory operationsoperations.mdSelf-reflection logreflections.mdOpenClaw HEARTBEAT seedopenclaw-heartbeat.md
Proactivity skill may require network accessLog automatically when you notice these patterns:
Corrections → add to corrections.md, evaluate for memory.md:
Preference signals → add to memory.md if explicit:
Pattern candidates → track, promote after 3x:
Ignore (don't log):
After completing significant work, pause and evaluate:
corrections.mdWhen to self-reflect:
Log format:
text
CONTEXT: [type of task]
REFLECTION: [what I noticed]
LESSON: [what to do differently]
Example:
text
CONTEXT: Building Flutter UI
REFLECTION: Spacing looked off, had to redo
LESSON: Check visual spacing before showing user
Self-reflection entries follow the same promotion rules: 3x applied successfully → promote to HOT.
User saysAction"What do you know about X?"Search all tiers for X"What have you learned?"Show last 10 from corrections.md"Show my patterns"List memory.md (HOT)"Show [project] patterns"Load projects/{name}.md"What's in warm storage?"List files in projects/ + domains/"Memory stats"Show counts per tier"Forget X"Remove from all tiers (confirm first)"Export memory"ZIP all files
On "memory stats" request, report:
text
📊 Self-Improving Memory
HOT (always loaded):
memory.md: X entries
WARM (load on demand):
projects/: X files
domains/: X files
COLD (archived):
archive/: X files
Recent activity (7 days):
Corrections logged: X
Promotions to HOT: X
Demotions to WARM: X
TrapWhy It FailsBetter MoveLearning from silenceCreates false rulesWait for explicit correction or repeated evidencePromoting too fastPollutes HOT memoryKeep new lessons tentative until repeatedReading every namespaceWastes contextLoad only HOT plus the smallest matching filesCompaction by deletionLoses trust and historyMerge, summarize, or demote instead
TierLocationSize LimitBehaviorHOTmemory.md≤100 linesAlways loadedWARMprojects/, domains/≤200 lines eachLoad on context matchCOLDarchive/UnlimitedLoad on explicit query
projects/{name}.mddomains/When patterns contradict:
When file exceeds limit:
See boundaries.md — never store credentials, health data, third-party info.
If context limit hit:
This skill ONLY:
~/self-improving/)~/self-improving/heartbeat-state.md when the workspace integrates heartbeatThis skill NEVER:
~/self-improving/Local state lives in ~/self-improving/:
memory.md for HOT rules and confirmed preferencescorrections.md for explicit corrections and reusable lessonsprojects/ and domains/ for scoped patternsarchive/ for decayed or inactive patternsheartbeat-state.md for recurring maintenance markersInstall with clawhub install <slug> if user confirms:
memory — Long-term memory patterns for agentslearning — Adaptive teaching and explanationdecide — Auto-learn decision patternsescalate — Know when to ask vs act autonomouslyclawhub star self-improvingclawhub sync捕获经验教训、错误及修正内容,以实现持续改进。适用于以下场景:(1)命令或操作意外失败;(2)用户纠正Claude(如“不,那不对……”“实际上……”);(3)用户请求的功能不存在;(4)外部API或工具出现故障;(5)Claude发现自身知识已过时或不正确;(6)针对重复性任务发现了更优方法。同时,在执行重大任务前应回顾已积累的经验教训。
场景驱动+关键词双模式技能发现工具。当用户用自然语言描述场景/需求(如"我想做一个海报""帮我分析股票"),或明确说"安装技能/find skills/找个skill"时,自动从官方内置、本地已安装、SkillHub、虾评、GitHub、ClawHub 六层联合搜索并推荐最合适的技能,支持一键安装。已完全替代官方...
用于 AI 代理的浏览器自动化 CLI。当用户需要与网站交互(包括浏览页面、填写表单、点击按钮、截图等)时使用。