memory-compaction
Pass
Audited by Gen Agent Trust Hub on Aug 22, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill is a documentation-oriented resource providing design patterns and pseudocode for managing LLM context windows. It does not contain functional code or scripts that execute on the host system.
- [INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by describing how to ingest and summarize conversation history.
- Ingestion points: The MemoryManager class in SKILL.md processes a history object that includes untrusted user messages.
- Boundary markers: The pseudocode uses delimiters ([Session memory: ... ]) to separate summarized history from the active context.
- Capability inventory: The skill does not provide or call any tools with high-risk capabilities such as network access, file system modification, or process execution.
- Sanitization: The logic focuses on summarization strategies; explicit sanitization of historical content is not detailed in the provided pseudocode.
Audit Metadata