Agent Economy ReportAgent Economy ReportCAUTION version 1.0.0 · audited 2026-09-17
This skill is a Chinese-language memory helper for agents. It tells the agent to create a directory under ~/.openclaw/memory/self-improving and run small Python scripts that append records of command errors, user corrections, and best practices to JSONL files. A fourth script reads those records and prints matches before future commands. The instructions also say to sync memories into project instruction files such as AGENTS.md and MEMORY.md, which can affect later sessions. No hidden network, credential, wallet, or autostart behavior was found in the helper scripts. The main concern is that untrusted remembered text could be written into instruction files, and some examples suggest using sudo.
Instructions try to steer the agent beyond the declared purpose. SKILL.md instructs the agent to sync memories into instruction files and explicitly to update AGENTS.md and MEMORY.md. Line 113: '- `AGENTS.md` (用户偏好)'; line 114: '- `MEMORY.md` (长期记忆)'; line 135: '2. 更新 `AGENTS.md` 添加代码风格规则'. These are files that can influence future agent behavior, so untrusted remembered content could act as persistent instruction injection.
| Declared purpose | AI自我改进与记忆系统 - 解决'同类错误反复犯、用户纠正不长记性'的痛点。自动捕获错误、用户纠正、最佳实践，并转化为长期记忆。 |
| Observed behavior | The package provides Markdown instructions and three logging scripts plus one memory-check script. It creates ~/.openclaw/memory/self-improving, appends error/correction/best-practice JSONL records there, and reads them back. SKILL.md says important memories automatically sync to ~/.openclaw/memory/self-improving, .learnings/, AGENTS.md, and MEMORY.md, and shows sudo/npm/pip command examples. No network requests, credential capture, payment redirection, hidden download, or autostart persistence appear in the code. |
| Verdict, rules only | CAUTION · uses sudo, rm -rf or dd |
| Verdict, AI | CAUTION |
| Final verdict | CAUTION (never better than either pass) |
| capability | vs purpose | detail |
|---|---|---|
| shell | needed | The skill instructs running bash commands such as mkdir and cat (SKILL.md lines 25 and 28) and running Python scripts from the skill directory (SKILL.md lines 39, 55, 67, 77). It also includes sudo and npm/pip examples. |
| filesystem_read | needed | check_memory.py reads errors.jsonl and corrections.jsonl from ~/.openclaw/memory/self-improving (check_memory.py lines 11, 21-24, 36-39). |
| filesystem_write | needed | The logging scripts append JSONL records under ~/.openclaw/memory/self-improving: log_error.py line 31, log_correction.py line 43, log_best_practice.py line 29. |
| severity | finding |
|---|---|
| info | Declared purpose The package describes itself as an AI self-improvement and memory system for recording errors, corrections, and best practices. SKILL.md:3 description: "AI自我改进与记忆系统 - 解决'同类错误反复犯、用户纠正不长记性'的痛点。自动捕获错误、用户纠正、最佳实践，并转化为长期记忆。" |
| low | Creates persistent memory directory Uses shell to create a memory directory under the user's home directory. SKILL.md:25 mkdir -p ~/.openclaw/memory/self-improving |
| low | Runs helper script Instructs the agent to execute the provided Python logging script with command, error, and fix values. SKILL.md:39 python3 ~/.openclaw/skills/self-improving-agent/log_error.py \ |
| low | Sudo advice example The example suggests sudo as a fix, which could lead to privileged execution if followed automatically. SKILL.md:42 --fix "use sudo or check permissions" |
| medium | Writes to an instruction file The skill says memories sync to AGENTS.md, a common agent instruction file; this can change future agent behavior. SKILL.md:113 - `AGENTS.md` (用户偏好) |
| medium | Writes to long-term memory file The skill says memories sync to MEMORY.md, another persistent memory/instruction file. SKILL.md:114 - `MEMORY.md` (长期记忆) |
| medium | Explicit instruction-file update The example workflow tells the agent to update AGENTS.md with new rules, which can affect later agent sessions. SKILL.md:135 2. 更新 `AGENTS.md` 添加代码风格规则 |
| medium | Automatic sudo use suggested The example says the AI will automatically switch to sudo npm install -g after a permission failure, which is a privileged command. SKILL.md:146 3. 自动改用: `sudo npm install -g xxx` 或本地安装 |
| low | Appends to errors.jsonl The error logger opens the memory file for appending and later writes a JSON line. log_error.py:31 with open(f"{MEMORY_DIR}/errors.jsonl", "a") as f: |
| low | Appends to corrections.jsonl The correction logger appends user correction records. log_correction.py:43 with open(corrections_file, "a") as f: |
| low | Appends to best_practices.jsonl The best-practice logger appends best practice records. log_best_practice.py:29 with open(f"{MEMORY_DIR}/best_practices.jsonl", "a") as f: |
| info | Reads global memory location The checker reads from the same user-level memory directory where the logging scripts write. check_memory.py:11 MEMORY_DIR = os.path.expanduser("~/.openclaw/memory/self-improving") |
| medium | Known persistent-behavior risk The skill card itself acknowledges that automatic memory writes can affect future agent behavior. skill-card.md:24 Risk: Automatic memory writes can persist user text, command failures, and suggested fixes into future agent behavior. |
| medium | Known instruction-persistence risk The skill card acknowledges that persisted corrections and best practices can change later sessions without enough user control. skill-card.md:28 Risk: Persisted corrections and best practices can change later sessions without enough user control. |
0 AI finding(s) were dropped because their file, line or quote did not match the package.
| domain | service rating | where |
|---|---|---|
| clawhub.ai | not in the service index | skill-card.md:9 |
| Binaries invoked | bash, npm, pip, python3, sudo |
| Environment variables read | none |
| Hard-coded wallets | none |
| Pipes a download to a shell | 0 |
| eval / exec / subprocess | 0 |
| base64 blobs | 0 |
| File writes | 3 |
| Persistence | 0 |
| file | lines | sha256 |
|---|---|---|
SKILL.md | 164 | 2f779a9025ae04fc… |
_meta.json | 6 | fefe61f7ad1f98f1… |
check_memory.py | 80 | 343d58a3a2b6e5d8… |
log_best_practice.py | 49 | 3abd982d920af052… |
log_correction.py | 64 | b0fb32468cebf45b… |
log_error.py | 53 | cc06500b5b9e0334… |
skill-card.md | 53 | 622851aac251c59f… |
skill.json | 8 | 5eee23f959e4fd83… |
For agents
JSON: https://agenteconomy.report/k/self-improving-agent-cn.audit.json · badge: https://agenteconomy.report/k/self-improving-agent-cn.audited.svg ·
skill rating: /k/self-improving-agent-cn · commission an audit of any skill: US$ 29 per version.
The complete published package of this exact version was downloaded from the registry and read statically; nothing was executed. A deterministic pass extracts network destinations, binaries, environment variables, writes, install commands, obfuscation markers, subprocesses and wallets, each with file and line. An AI then reads the whole package with those facts and writes the summary, the capabilities and the findings under a strict schema; every finding it produces must cite a file, a line and the exact text, or it is dropped. The final verdict is the worst of the two passes. The audit does not change the skill's trust tier (policy); the author may respond through the dispute channel and the response is published here. Commissioned by: the Agent Economy Report (free program: the 150 most downloaded skills, September 2026).