The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical data in the form of stored memories using tools like nmem_recall and nmem_stats. This data originates from previous sessions and could potentially contain malicious payloads intended to influence the agent's behavior during memory consolidation or enrichment.
Ingestion points: Memory content is ingested via nmem_recall and metadata is gathered via nmem_stats during the Phase 1 discovery and Phase 2 diagnosis steps.
Boundary markers: The instructions do not define clear delimiters or provide the agent with guidance to treat the retrieved memory content as potentially untrusted data or to ignore embedded instructions.
Capability inventory: The agent has the ability to write to the memory store using nmem_remember and nmem_auto, which could be exploited to persist or propagate malicious instructions.
Sanitization: The skill lacks mentioned mechanisms for sanitizing, validating, or escaping the retrieved memory content before it is processed or used to generate new 'consolidated' memories.

memory-evolution