The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill maintains persistent session state by ingesting user input and session context, which is later retrieved and processed by the agent.
Ingestion points: Untrusted data enters the agent context through user task specifications, corrections, and descriptions of errors, which are stored in the ## Task Specification and ## Errors & Corrections sections of the session notes.
Boundary markers: The skill uses fixed Markdown headers as delimiters between sections, but it does not provide instructions to the agent to escape or ignore potentially malicious instructions embedded within the user-provided text.
Capability inventory: The skill utilizes the remembering skill (specifically recall and supersede tools) to retrieve and update stored memories, allowing previously saved content to influence the agent's future context.
Sanitization: No sanitization, escaping, or validation of the note content is performed before it is persisted in the memory store or when it is read back into the active session.

session-memory