/install amitpnyc-self-improving-agent
Self-Improving Agent
Capture useful lessons while context is fresh. Keep the default behavior lightweight, private, and append-only.
Core rules
- Prefer signal over volume.
- Default to logging, not self-modifying instructions.
- Log only meaningful corrections, failures, recurring issues, durable conventions, or explicit “remember this” requests.
- Never log secrets, tokens, private keys, environment variables, raw customer data, or full sensitive transcripts unless the user explicitly asks.
- Prefer short summaries and redacted excerpts over raw output.
- In OpenClaw, treat edits to
AGENTS.md,SOUL.md,TOOLS.md, andMEMORY.mdas high-authority changes. Do not make them automatically unless the user asked or the workspace explicitly authorizes it. - In OpenClaw, update
MEMORY.mdonly in trusted direct or main-session contexts, not by default from shared/group contexts or routine subagent work.
Initialize once
Store the skill in \x3Cworkspace>/skills/self-improving-agent/ for workspace-local use or ~/.openclaw/skills/self-improving-agent/ for shared use.
Create .learnings/ in the active workspace root or project root and ensure these files exist:
.learnings/LEARNINGS.md.learnings/ERRORS.md.learnings/FEATURE_REQUESTS.md
Never overwrite an existing log file.
Where to log
ERRORS.md— command failures, tool crashes, API/integration issues, reproducible environment problemsLEARNINGS.md— user corrections, outdated assumptions, best practices, non-obvious debugging conclusions, recurring workflow hardeningFEATURE_REQUESTS.md— capabilities the user wanted but the system or workflow did not support
Use these learning categories when relevant:
correctioninsightknowledge_gapbest_practice
Default operating mode
Use append-only capture by default:
- append a short structured entry to the right file
- link related prior entries when issues recur
- suggest promotion when warranted
- do not automatically edit long-lived instruction or memory files unless authorized
Promotion rules
Promote a learning only when it is repeated, durable, broad, costly to forget, or explicitly marked permanent by the user.
Do not promote one-off incidents, transient outages, machine-specific glitches, speculative opinions, or unclear temporary preferences.
Promote to the smallest durable home:
AGENTS.md— workflow rules and execution guidanceTOOLS.md— tool gotchas and environment notesSOUL.md— behavioral principles and communication styleMEMORY.md— durable user/project facts- project instruction files such as
CLAUDE.mdor.github/copilot-instructions.md— only when the learning is project-wide
When promoting, distill the lesson into a short rule. Do not copy the full log entry.
Dedupe
Before adding a new entry:
- scan for a related item
- use
See Alsofor related entries - prefer updating recurrence metadata over creating near-duplicates
Use a stable Pattern-Key for repeated workflow issues when helpful.
Minimal entry shapes
## [LRN-YYYYMMDD-XXX] category
**Logged**: ISO-8601 timestamp
**Priority**: low | medium | high | critical
**Status**: pending
**Area**: frontend | backend | infra | tests | docs | config | workflow
### Summary
One-line lesson
### Details
What happened and what is correct now
### Metadata
- Source: conversation | error | user_feedback | investigation
- Related Files: path/to/file.ext
- See Also: LRN-YYYYMMDD-XXX
- Pattern-Key: optional-key
- Recurrence-Count: 1
## [ERR-YYYYMMDD-XXX] system_or_command
**Logged**: ISO-8601 timestamp
**Priority**: medium | high | critical
**Status**: pending
**Area**: frontend | backend | infra | tests | docs | config | workflow
### Summary
What failed
### Error
Short error text or redacted excerpt
### Context
- Operation attempted
- Relevant inputs or parameters
### Metadata
- Reproducible: yes | no | unknown
- Related Files: path/to/file.ext
- See Also: ERR-YYYYMMDD-XXX
## [FEAT-YYYYMMDD-XXX] capability-name
**Logged**: ISO-8601 timestamp
**Priority**: low | medium | high
**Status**: pending
**Area**: frontend | backend | infra | tests | docs | config | workflow
### Requested Capability
What the user wanted
### User Context
Why it mattered
### Metadata
- Frequency: first_time | recurring
- Related Features: feature-or-workflow
Cross-session sharing
Share learnings across sessions only when the user wants that behavior or the environment explicitly supports it.
When sharing:
- send a short sanitized summary
- include only needed IDs and file paths
- do not send raw transcripts or secret-bearing output by default
- in OpenClaw, prefer
sessions_sendsummaries over transcript forwarding
OpenClaw notes
- Prefer
openclaw skills install \x3Cslug>for installation guidance. - Do not include Claude/Codex-style hook examples unless you are shipping a real OpenClaw hook.
- If you add hook automation, ship
HOOK.md+handler.tsand enable it withopenclaw hooks enable \x3Cname>. - If the skill later references bundled files, use
{baseDir}. - Keep setup instructions cross-platform.
Default behavior
When something notable happens:
- Decide whether it clears the logging threshold.
- Append a concise entry.
- Link related prior entries if applicable.
- Suggest promotion only if the lesson is repeated or durable.
- Promote sparingly and in distilled form.
If unsure whether something deserves promotion, keep it in .learnings/ and move on.
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install amitpnyc-self-improving-agent - 安装完成后,直接呼叫该 Skill 的名称或使用
/amitpnyc-self-improving-agent触发 - 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
Self-Improving Agent 是什么?
Log high-signal corrections, tool failures, feature requests, and recurring workflow lessons to a lightweight .learnings/ directory, then promote only repeat... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 74 次。
如何安装 Self-Improving Agent?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install amitpnyc-self-improving-agent」即可一键安装,无需额外配置。
Self-Improving Agent 是免费的吗?
是的,Self-Improving Agent 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
Self-Improving Agent 支持哪些平台?
Self-Improving Agent 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。
谁开发了 Self-Improving Agent?
由 AmitPNYC(@amitpnyc)开发并维护,当前版本 v1.0.0。