The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user conversation history to create rolling summaries and extract entities into semantic storage. This ingestion surface is susceptible to indirect prompt injection where an attacker might try to influence the summary or persistent memory.
Ingestion points: Conversation history (last 10 messages) and project files identified via grep.
Boundary markers: The skill includes explicit instructions for 'Amnesia Exploitation Defense' and 'Context Bleed Prevention' to safeguard against such attacks.
Capability inventory: Reads message history and file system; writes to AUDIT_LOG.md, VOTING_LOG.md, and the semantic/ directory; deletes (unlinks) temporary files.
Sanitization: The 'Sanitized Forget Protocol' specifically instructs the agent to clear cached credentials and PII before summarizing content.
[DATA_EXPOSURE]: The skill manages an audit trail and semantic memory. It defines strict rules for AUDIT_LOG.md and VOTING_LOG.md, treating them as immutable write-only ledgers. It also includes specific instructions to scrub credentials and PII from context windows during transitions.
[COMMAND_EXECUTION]: The skill utilizes standard system tools like grep for file identification and standard filesystem operations for the 'Forget' protocol (unlinking files). These are used within the scope of intended memory management tasks and do not represent arbitrary command execution.

memory-optimization