The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions recommend installing an external tool via npx claude-mem install and a plugin from a third-party repository thedotmack/claude-mem. These resources are hosted on public registries and are not associated with trusted organizations, posing a risk of downloading unverified code.
[REMOTE_CODE_EXECUTION]: The use of npx to install and execute the claude-mem package involves fetching code from the NPM registry and running it on the local system. Without a verified or trusted publisher, this constitutes remote code execution from an unverifiable source.
[INDIRECT_PROMPT_INJECTION]: The core functionality of the skill—compressing and re-injecting conversation history and project context—presents an indirect prompt injection surface. Malicious instructions contained within the analyzed transcripts or project files could be stored in the memory snapshots and later influence the agent's behavior when the memory is loaded.
Ingestion points: Conversation transcripts and project context extracted during the /mem compress command as described in SKILL.md.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions to isolate historical memory from current session instructions.
Capability inventory: The skill interacts with the local filesystem (.claude-mem/ directory) to read and write memory snapshots.
Sanitization: There is no evidence of sanitization or filtering of the content being compressed into memory.

claude-mem