Pgmemory

Persistent semantic memory for OpenClaw agents — PostgreSQL + pgvector

This skill appears to implement a coherent PostgreSQL+pgvector memory system, but take these precautions before installing/ running the setup wizard: 1) Inspect scripts/setup.py and confirm you are comfortable with it modifying AGENTS.md, writing a config to ~/.openclaw/pgmemory.json, and installing/configuring cron/docker. 2) Decide how to supply embedding API keys: the wizard can store a key in plain JSON under your home directory (CHANGELOG explicitly added this behavior) — consider using a local provider (Ollama) or an env var and restrict file permissions if you allow config storage. 3) Backup AGENTS.md and any agent workspaces before running --sync-agents or the automatic injection. 4) The requirements mention numpy but requirements.txt only lists psycopg2-binary; ensure you install necessary Python deps manually. 5) For least privilege, run the setup in a sandboxed environment (or test agent) first and review network calls (voyage/openai endpoints) if you must avoid sending embedding data externally. If you want, I can point out the exact lines in setup.py that modify AGENTS.md, create cron jobs, or persist the api_key so you can review/patch them.

Type: OpenClaw Skill Name: pgmemory Version: 1.2.0 The skill is classified as suspicious primarily due to a significant Remote Code Execution (RCE) vulnerability in `scripts/setup.py`. The script uses `curl -fsSL https://get.docker.com | sh` to install Docker, which executes arbitrary code downloaded from the internet without prior validation, posing a supply chain risk if the source domain were compromised. Additionally, `scripts/setup.py` performs high-privilege system modifications (e.g., `sudo usermod`, `crontab` scheduling) and modifies `AGENTS.md` to inject instructions for the OpenClaw agent, which, while intended for legitimate skill integration, demonstrates the use of prompt injection as an attack surface.