Yaoyao Memory V2

What to consider before installing: - Metadata mismatch: The registry header claims no required env vars, but SKILL.md and _meta.json require LLM/EMBEDDING API keys and llm_config.json. Ask the publisher to clarify which credentials/paths are truly required. - Secrets exposure risk: The package reads/writes ~/.openclaw/credentials/secrets.env (a global credentials file). Prefer creating dedicated, scoped API keys and store them in a per-skill config (llm_config.json) rather than allowing the skill access to global secrets.env. Review/limit file permissions. - Silent recording: The skill is explicitly '静默自动' (silent automatic). That means it will persistently record conversational data across sessions without prompting. If you have sensitive conversations, do not enable this until you audit the code and configuration. - Prompt-injection indicators: The SKILL.md contains patterns flagged as prompt-injection attempts. Treat SKILL.md and bundled instructions as untrusted until reviewed. Verify the agent’s policy/guardrails prevent skills from overriding system prompts. - Network endpoints: The code posts to configurable LLM/embedding endpoints using provided API keys. Ensure you use limited-scope/test keys and monitor token usage; consider running with an on-prem or sandbox LLM endpoint first. - Run in sandbox first: Install and run the skill in an isolated environment (VM/container) and use test API keys. Inspect runtime behavior (which files are created/modified, outbound connections, logs) before granting access to real credentials or production data. - Review or restrict files: Manually review scripts that touch ~/.openclaw/credentials and any code that reads/writes memory files. If you proceed, prefer configuring the skill to use a dedicated memory directory and dedicated credential files, and disable IMA cloud sync unless you trust that endpoint. If you want, I can: - Highlight specific files/functions that read credentials or perform network calls so you can focus your code review; - Suggest a minimal safe configuration (paths and scoping) to test the skill in a sandbox; - Produce commands to search the bundle for other suspicious patterns (hard-coded endpoints, exec/system calls, or telemetry).