The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a mechanism to save user-supplied information into a persistent MEMORY.md file. Because this file is designed to be read by the agent at the start of every session, it creates a surface for instructions to be persisted and subsequently followed by the agent in a different context.
Ingestion points: User input parsed and stored via the /si:remember command described in SKILL.md.
Boundary markers: No delimiters or explicit instructions to ignore embedded commands are used when writing the entry to MEMORY.md.
Capability inventory: The skill has the ability to append to local files (MEMORY.md) and execute shell commands (grep, sed) to manage this data.
Sanitization: There is no evidence of sanitization or validation of the input string before it is appended to the memory file.
[COMMAND_EXECUTION]: The skill uses shell commands to check for duplicate entries and determine the correct path for memory storage.
The duplicate check in Step 2 uses grep on keywords derived from user input. If the platform does not properly escape these keywords before passing them to the shell, it could lead to argument injection or command manipulation.
The skill uses pwd and sed to dynamically construct the path to the memory directory.

remember