ABot-Recon is a streaming 3D reconstruction system from AMAP CVLab, the computer vision research lab of Alibaba's AMAP mapping and navigation service, built around the idea that long-horizon reconstruction does not require accumulating long-range learned memory. At each timestep the model caches key-value features from the preceding 11 frames, predicts a point map for the current frame in its own camera coordinate system, estimates the relative pose to the previous frame, and recovers the full trajectory and point cloud by composing these adjacent poses sequentially; a motion-visual rotation refiner and a composition-aware pose loss limit drift as poses accumulate over long sequences. Because per-frame state and compute stay independent of total sequence length, the demo script supports streams up to 22,000 frames by default. The README reports that camera-pose tracking on Oxford Spires reaches an ATE of 4.35 m and RPE-R of 0.12 degrees with the streaming model alone and no loop closure, dense reconstruction on the same scene reaches a Chamfer Distance of 1.37 m and F1 of 91.81% at a 4 m threshold, and streaming throughput on KITTI-02 at 504x280 resolution is 24.45 FPS using 6.71 GiB on an NVIDIA H100. An optional loop-closure backend retrieves candidate revisit pairs with DINOv2-SALAD descriptors, predicts relative-pose constraints with ABot-Recon itself, and refines the trajectory through sparse pose-graph optimization, while the base model runs with no loop-closure dependency at all. A public checkpoint is hosted on Hugging Face and ModelScope and downloads automatically through the Python API or the demo.py CLI, and an interactive Hugging Face Space demo has been live since August 31, 2026. The project's own release-status checklist marks the public checkpoint and inference/evaluation code as shipped while training code and recipes remain unreleased, planned for September 30. Source code is Apache-2.0 licensed, but the model weights are governed by a separate MODEL_LICENSE.md, and the authors ask users to review a separate model usage guidelines document before use. The work builds on the Pi3 architecture and credits CroCo, DUSt3R, DINOv2, SALAD and FlashInfer as influences.