memory-systems
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes patterns for storing and retrieving user-provided facts and entity relationships. Because these systems ingest and persist external data which is subsequently retrieved into the agent's context, they create a surface for indirect prompt injection attacks.
- Ingestion points: Pattern implementations in
SKILL.md(e.g.,store_fact,remember_entity). - Boundary markers: Not present in the provided examples; retrieval logic does not show delimiters or warnings to ignore instructions within stored memories.
- Capability inventory: The described patterns involve file system writes (
save_json) and interactions with databases (vector stores, graph databases). - Sanitization: The provided snippets do not include input validation or sanitization for the facts being stored.
Audit Metadata