Remembering Conversations

Pass

Audited by Gen Agent Trust Hub on Feb 27, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The install-hook script establishes a lifecycle hook at ~/.claude/hooks/sessionEnd to automate the indexing of conversation data at the end of each session.
  • [DATA_EXFILTRATION]: Conversation content is read from local logs and sent to the Anthropic API via the official SDK to generate summaries for the search index.
  • [EXTERNAL_DOWNLOADS]: The skill downloads the all-MiniLM-L6-v2 embedding model from Hugging Face during the initial configuration.
  • [PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes untrusted historical conversation data. Ingestion points: ~/.claude/projects/*.jsonl files containing previous user and assistant messages. Boundary markers: The subagent template (search-agent.md) provides structural headers but lacks robust delimiters for the ingested data. Capability inventory: File system access (read/write), network API calls (summarization), and CLI tool execution. Sanitization: Conversation content is summarized and indexed without explicit sanitization or filtering.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 27, 2026, 03:35 PM