Once Screenpipe has months of screen and audio indexed locally, what stops a compromised or over-permissioned agent from querying the whole history through the port 3030 API instead of just the current task's context? Continuous recording solves the memory problem, but it turns "give the agent access to my second brain" into a much bigger blast radius than giving it access to one document. Is there per-query scoping or a time-window restriction on what an agent can pull back, or is authentication the only gate?