The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted dialogue data from users which is subsequently compressed into atomic facts and used to generate responses. An attacker could potentially embed malicious instructions within dialogues that, when retrieved, could influence agent behavior.
Ingestion points: The memory_add tool and add_dialogue API method (SKILL.md, references/api-reference.md).
Boundary markers: The instructions do not define explicit boundary markers or 'ignore' instructions for the dialogue content.
Capability inventory: The system uses LLMs for fact extraction, coreference resolution, and answer generation (references/architecture.md).
Sanitization: The skill mentions 'density gating' and 'reformulation' which may reduce noise, but does not specify adversarial filtering mechanisms.
[CREDENTIALS_UNSAFE]: The documentation for the cloud-hosted MCP server instructs users to visit https://mcp.simplemem.cloud and enter their OpenRouter API key to generate a secondary authentication token. This practice involves sharing a high-privilege secret with a third-party infrastructure provider.
[EXTERNAL_DOWNLOADS]: The skill relies on external code execution through the installation of the simplemem package via pip and recommends cloning the source code from the aiming-lab/SimpleMem GitHub repository.

simplemem