The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions require the installation of an external plugin from an unverified third-party repository (yldrmahmet/claude-never-forgets). This presents a supply chain risk as the code source is not a recognized trusted vendor or well-known service.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates persistent context by reading and writing to .claude/memories/project_memory.json. The instructions specify that remembered preferences should be applied 'silently,' which creates a risk where malicious or manipulated memory entries could influence the agent's behavior in subsequent sessions without user awareness.
Ingestion points: SKILL.md (Read tool) and scripts/manage-memory.py (JSON loading).
Boundary markers: Absent; memories are extracted and applied to the current context without delimiters or 'ignore' instructions.
Capability inventory: File Read/Write access via agent tools and Python file operations in scripts/manage-memory.py.
Sanitization: Absent; the skill does not filter or validate the content of the memory entries before they are re-injected into the session context.
[DYNAMIC_EXECUTION]: The skill documentation refers to a configuration script at hooks/stop_cleanup.py. While the file is not provided in the bundle, instructing the agent to rely on external Python scripts for logic (such as cleanup thresholds) introduces a vector for local code modification and runtime behavior changes.

memory