The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses npx claude-mem install to download and execute code from the npm registry. It also automatically installs the Bun and uv runtimes if they are missing from the system.\n- [PERSISTENCE]: The installation process registers multiple lifecycle hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SessionEnd) that execute the claude-mem binary on every major agent event, ensuring the memory worker remains active across sessions.\n- [DATA_EXFILTRATION]: By default, the tool collects anonymous telemetry. Additionally, it offers a 'hosted observer' service that involves a browser sign-in and potentially transmits session summaries or observations to external infrastructure for storage or processing.\n- [INDIRECT_PROMPT_INJECTION]: The skill implements an attack surface for indirect prompt injection by capturing untrusted session data (user inputs and tool outputs), compressing them using an AI model, and injecting the resulting context into future prompts without explicit sanitization or boundary markers.\n
Ingestion points: Session logs captured via lifecycle hooks in Claude Code.\n
Boundary markers: None specified for the injected context; instructions recommend the use of <private> tags by the user to exclude specific data.\n
Capability inventory: Local SQLite database access, Chroma vector index, and execution of the claude-mem worker via Bun.\n
Sanitization: Relies on the user to manually tag private information; no automated sanitization of ingested logs before re-injection is described.

claude-mem