The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled Python script (memory_preflight.py) to check the health and status of the local Claude Mem service. The analysis of this script confirms it is restricted to read-only operations, using mode=ro and immutable=1 for SQLite connections to prevent database corruption or unauthorized modification.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses local SQLite databases related to Claude Mem (claude-mem.db). However, the access is strictly read-only for diagnostic purposes (checking row counts and queue depths) and does not include any patterns for exfiltrating this data to external servers. The diagnostic results are returned locally to the agent to inform task planning.
[REMOTE_CODE_EXECUTION]: There is no remote code execution detected. The skill uses a local script for preflight checks and relies on established local tools like rg (ripgrep) for code exploration.
[PROMPT_INJECTION]: The instructions contain strict boundary markers and freshness rules (e.g., 'Do not present memory-derived facts as current without verification'). It explicitly instructs the agent to treat historical data as hypotheses until verified against current code, which is a defensive measure against accidental reliance on stale or potentially injected historical context.
[SAFE]: The skill follows least-privilege principles, explicitly prohibiting administrative actions on memory services (e.g., 'Do not start, stop, restart, kill, or reconfigure memory workers').

gpt-engineer-mem