The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents an interface for persistent data storage that can be exploited for indirect prompt injection.
Ingestion points: Tools like memory_save, memory_lesson_save, and memory_slot_append in REFERENCE.md ingest raw text content into long-term memory.
Boundary markers: No explicit boundary markers or instructions to isolate stored data from agent instructions are described in the documentation.
Capability inventory: The skill provides a reference for 54 tools, including those for file system manipulation, inter-agent messaging, and graph-based retrieval.
Sanitization: While SKILL.md mentions that tool handlers whitelist fields, there is no mention of content-level sanitization to prevent the storage of malicious instructions.
[DATA_EXFILTRATION]: The reference includes tools capable of moving data across environment boundaries.
memory_export provides a mechanism for bulk data export in JSON format.
memory_signal_send enables transmission of data between agents or to broadcast channels.
memory_mesh_sync allows for memory state synchronization with peer agent instances.
[COMMAND_EXECUTION]: The tool index includes capabilities that interface with external system states and development workflows.
memory_commit_lookup and memory_commits interact with Git repository state and metadata.
memory_checkpoint allows agent actions to be gated by external events such as CI/CD results.

agentmemory-mcp-tools