ag2-knowledge-and-memory

Pass

Audited by Gen Agent Trust Hub on May 14, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill provides documentation and runnable examples for implementing persistent memory and history management using the ag2 framework's built-in APIs. The use of local storage in assets/journal_companion.py is safely handled within temporary system directories.\n- [INDIRECT_PROMPT_INJECTION]: The skill implements features for injecting stored historical context into active agent prompts, which creates a potential surface for indirect prompt injection if the stored data originates from untrusted sources.\n
  • Ingestion points: WorkingMemoryPolicy and EpisodicMemoryPolicy read data from a KnowledgeStore (e.g., /memory/working.md and /memory/conversations/) to augment the LLM's context window.\n
  • Boundary markers: Boundary markers for injected context are not explicitly defined in the provided examples; the skill relies on the framework's internal event handling and agent instructions.\n
  • Capability inventory: The agents configured in the examples have the capability to write to the local filesystem via KnowledgeStore and perform LLM-driven summarization of event history.\n
  • Sanitization: The examples utilize ConversationPolicy to filter non-dialogue events (like system logs) from the prompt, though explicit content sanitization of user-provided entries is not demonstrated.
Audit Metadata
Risk Level
SAFE
Analyzed
May 14, 2026, 01:22 AM
Security Audit — agent-trust-hub — ag2-knowledge-and-memory