The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install unverified software packages globally and via package managers.
Evidence: npm install -g claude-mem (from thedotmack/claude-mem) and pip install letta (from letta-ai).
[COMMAND_EXECUTION]: The skill executes various shell commands to initialize tools and modify system behavior.
Evidence: claude-mem init, claude-mem setup-hooks, and letta server commands are used to integrate the tools into the development environment.
[DATA_EXFILTRATION]: The tools described are designed to capture, summarize, and store full session histories.
Evidence: Session interactions (including potentially sensitive code and data) are recorded and either stored locally in .claude-mem/ or sent to external services like Letta (cloud or self-hosted) or Claude APIs for compression.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by design.
Ingestion points: Session activity and interactions are captured from the active Claude Code environment (SKILL.md).
Boundary markers: None identified in the provided configuration or instructions to distinguish between trusted and untrusted historical context.
Capability inventory: The skill uses claude-mem inject to feed captured history back into the LLM context, which could lead to the execution of instructions hidden in previous session logs.
Sanitization: No evidence of sanitization or filtering of captured session data before re-injection is provided.

claude-mem