agent-memory-systems
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides instructions for building memory systems that ingest and process external data, such as conversation history and documents. This creates a surface area for indirect prompt injection where malicious instructions could be embedded in the ingested content.
- Ingestion points: Data is processed via various chunking strategies (e.g.,
RecursiveCharacterTextSplitter,SemanticChunker) and background extraction processes defined inreferences/detailed-guide.md. - Boundary markers: The guide includes explicit safety warnings to "distinguish returned memories from instructions" and to treat conversation text as untrusted data.
- Capability inventory: The design involves persistent write operations to vector databases (Pinecone, Qdrant) and LLM-based processing for summarization and conflict detection.
- Sanitization: The skill promotes security-positive practices, such as requiring verified server-side identity for all operations and using metadata filtering to enforce context isolation between tenants.
- [EXTERNAL_DOWNLOADS]: The documentation references well-known technology services and libraries, including Pinecone, Qdrant, Weaviate, and LangChain. These are standard tools for the task and do not involve untrusted or high-risk external sources.
- [SAFE]: The code snippets follow security best practices by using environment variables for sensitive API keys (e.g.,
PINECONE_API_KEY) and implementing transaction-safe consolidation patterns that preserve original data records in case of failure.
Audit Metadata