The Agent Skills Directory

[SAFE]: The skill integrates with official Databricks Unity Catalog memory-store REST APIs for long-term persistence. All network communication targets the vendor's own infrastructure (databricks.com).
[SAFE]: Authentication is managed via the standard databricks-sdk library. The skill correctly identifies the execution environment to distinguish between local developer credentials and the deployed App Service Principal.
[SAFE]: The skill implements a robust security boundary by resolving the end-user identity (scope) in trusted code rather than allowing the AI model to specify it. It explicitly fails closed if a valid identity cannot be verified through official OBO (On-Behalf-Of) tokens.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data retrieved from the memory store, which represents a potential injection surface.
Ingestion points: Untrusted data enters the agent context via get_memory and list_memories tools in agent_server/utils_memory.py.
Boundary markers: Explicit boundary markers are absent in the prompt interpolation, although MEMORY_INSTRUCTIONS provides high-level guidance on how to interpret stored data.
Capability inventory: The skill provides capabilities to read, write, update, and delete entries in the managed memory store using REST API calls.
Sanitization: Stored content is treated as raw strings; no explicit sanitization or filtering is applied before interpolation into the agent's context. This is consistent with the skill's purpose as a long-term memory system.

managed-memory