agent-memory-patterns
Pass
Audited by Gen Agent Trust Hub on Oct 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill consists of documentation and reference implementations for agent memory management. The provided code snippets demonstrate standard architectural patterns for context budgeting, working memory data structures, and vector-based retrieval.\n- [INDIRECT_PROMPT_INJECTION]: The skill describes systems designed to ingest and process untrusted data from user interactions and external knowledge sources. While this establishes an indirect prompt injection attack surface, the content focuses on design patterns to manage this surface (e.g., memory compression, consolidation, and tiered visibility) rather than introducing exploitable vulnerabilities.\n
- Ingestion points: SKILL.md outlines patterns for capturing session history, event logs, and distilled facts into persistent memory stores.\n
- Boundary markers: The design patterns recommend visibility tiers and namespacing, though the illustrative code snippets prioritize simplicity over comprehensive boundary enforcement.\n
- Capability inventory: The patterns involve interactions with vector databases and LLM inference providers.\n
- Sanitization: The documentation identifies the risk of stale or poisoned memory and suggests freshness checks and consolidation as mitigation strategies.
Audit Metadata