This work presents an on-chip memory subsystem envisioned for real-time systems performing Oriented FAST and Rotated BRIEF (ORB) feature extraction in Simultaneous-Localization and Mapping (SLAM) systems. For autonomous navigation of battery-powered devices, feature-based SLAM is a computationally frugal alternative to direct methods. This paper thoroughly analyses ORB multiple memory access patterns, exploring parallelism and hardware-centric algorithmic enhancements, reducing requirements on bandwidth and redundant accesses. Enabling those, a suitable multi-bank parallel memory featuring run-time reconfigurable address generation, image allotment, and close-to-memory data-shuffling is proposed. As case study, a 30 frames-per-second (FPS) VGA-resolution ORB-capable 8-bank memory is evaluated using 22 FDx technology.