The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The memory-management.md file explicitly identifies 'API endpoints and credentials' as key information types to preserve within the knowledge graph using the mcp__memory__add_observations tool. Storing secrets in a shared memory graph increases the risk of exposure to other agents or sessions accessing the same memory space.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture is centered on capturing and redistributing context across sessions and agents, which creates a significant surface for indirect prompt injection if the ingested data contains malicious instructions.
Ingestion points: Untrusted data is ingested from the knowledge graph through mcp__memory__search_nodes and mcp__memory__open_nodes, and from filesystem artifacts in .claude/ and /tmp/ directories as described in references/patterns.md.
Boundary markers: The skill lacks instructions for using delimiters or explicit boundary markers when re-injecting stored context into an agent's active prompt.
Capability inventory: The skill has the capability to write to the local filesystem (session context files) and perform CRUD operations on the MCP memory server.
Sanitization: There is no mechanism described for sanitizing, escaping, or validating stored observations before they are used to influence agent behavior in subsequent sessions.

context-management