The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a significant ingestion surface for indirect prompt injection as it mines project source files and Claude Code conversation transcripts into a persistent database.\n
Ingestion points: The mempalace mine and mempalace sweep commands ingest data from project directories and ~/.claude/projects/ transcripts.\n
Boundary markers: While the skill documentation mentions sanitize_content for storage, there are no explicit instructions for the agent to use delimiters or ignore instructions embedded in retrieved memory chunks.\n
Capability inventory: The skill provides an MCP server with 29 tools for database interaction and installs shell hooks for automatic execution.\n
Sanitization: Reference to sanitize_name() and sanitize_content() functions in mempalace/config.py indicates some level of data validation.\n- [PERSISTENCE]: The skill provides a mempalace hook install command that configures shell scripts (mempal_save_hook.sh and mempal_precompact_hook.sh) to run automatically when the agent stops or compacts context. This ensures session data is persisted across executions.\n- [EXTERNAL_DOWNLOADS]: The instructions direct the user to install the mempalace package using uv, pipx, or pip. While the package is authored by the skill's creator, the installation commands provided do not pin the package to a specific version.\n- [COMMAND_EXECUTION]: The skill utilizes several CLI commands and shell scripts for operation, mining, and maintenance, including direct path access to local directories for reading and writing data.

mempalace