folded-memory-implementation

Pass

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The memory compaction and distillation logic is vulnerable to indirect prompt injection from conversation history.
  • Ingestion points: Raw user conversation turns are ingested by the compact_l1_to_episode and distill_episodes_to_semantic functions in SKILL.md.
  • Boundary markers: The prompt templates EPISODE_COMPACTOR_PROMPT and SEMANTIC_EXTRACTOR_PROMPT do not utilize delimiters or specific instructions to ignore malicious directives embedded in the text being summarized.
  • Capability inventory: The skill provides logic to persist this potentially poisoned memory to the local filesystem via save_folded_memory in SKILL.md.
  • Sanitization: There is no evidence of sanitization or filtering of the conversation turns before they are processed by the LLM for memory folding.
  • [EXTERNAL_DOWNLOADS]: The README.md file contains instructions to install the skill using npx skills add, which fetches content from the author's repository.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 6, 2026, 01:34 AM
Security Audit — agent-trust-hub — folded-memory-implementation