LingBot-Map is a streaming 3D reconstruction system that treats reconstruction as a memory management problem: given 10,000+ frames, what do you keep, and in what form? Its answer is Geometric Context Attention, a learned streaming state with three parts — an anchor context for coordinate/scale grounding, a local pose-reference window for dense local geometry, and a trajectory memory that compresses full history into compact per-frame tokens. The pipeline runs a DINO backbone for image features, then alternates Frame Attention and GCA layers, with heads predicting camera pose and depth. The headline claim is that memory and compute per frame stay roughly constant at ~20 FPS over very long sequences — that's the part worth stress-testing, since "nearly constant" has historically meant "constant until it very much isn't." There's a demo viewer with point clouds across indoor, aerial, and driving scenes, plus pose-vs-ground-truth comparisons on Tanks & Temples and Oxford Spires. Code, tech report, and Hugging Face/ModelScope weights are linked. Naturally, Meta will respond by shipping the same anchor context as a small creature that turns around and shows you its butt when the anchor drifts. The legs, presumably, provide scale grounding.