The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/check_hermes_ebbinghaus.py performs dynamic code loading by modifying sys.path with a path provided via the --repo command-line argument. It subsequently imports and executes code from the plugins.memory package within that directory. This allows for arbitrary code execution if the agent is directed to run the script against an untrusted or compromised directory.
[COMMAND_EXECUTION]: The SKILL.md instructions direct the agent to execute shell commands and Python scripts using local environment variables (e.g., $env:USERPROFILE) and user-supplied repository paths. This includes running pytest and diagnostic scripts that interact with the local filesystem and external repository code.
[DATA_EXPOSURE]: The skill is designed to read local configuration files (config.yaml) and access SQLite database files (ebbinghaus_memory.db) within the $HERMES_HOME directory. While these are required for the skill's stated purpose of managing the Hermes memory provider, they represent access to potentially sensitive application data.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text as 'memories' for storage and retrieval. This creates an attack surface where malicious instructions embedded in stored data could influence the agent's behavior during the 'recall' or 'rehearse' actions.
Ingestion points: External text stored via the remember action in SKILL.md and scripts/check_hermes_ebbinghaus.py.
Boundary markers: None identified in the provided schema to distinguish between data and instructions.
Capability inventory: File system access (SQLite), shell command execution (pytest), and dynamic Python code loading.
Sanitization: No explicit sanitization or validation of the memory content is performed before processing.

human-memory-emulation