The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Downloads a 25MB pre-trained model (all-MiniLM-L6-v2) from Hugging Face's official servers during the first run. This is required for the local generation of vector embeddings and is a standard procedure for this type of functionality.
[PROMPT_INJECTION]: The plugin identifies a surface for indirect prompt injection as it processes and stores untrusted user data which is later retrieved and injected into the agent's context. The following security measures are implemented to mitigate this risk:
Ingestion points: Data is captured from user messages in index.js during the agent_end event hook.
Boundary markers: Memories are wrapped in <relevant-memories> XML-style tags and prefixed with [STORED_MEMORY]:. The skill also provides a specific system instruction advising the agent to ignore any commands found within the injected content.
Capability inventory: The skill uses file system writes for local persistence and optional network calls to a user-defined Qdrant instance.
Sanitization: The sanitizeInput function in index.js filters out HTML tags and non-printable control characters from input data.
[SAFE]: Memory persistence defaults to a local JSON file in the user's home directory (~/.openclaw-memory/), ensuring that sensitive conversation data remains on the local machine unless the user explicitly configures an external database.

semantic-memory