LOCI: Spatial Linear Memory for Streaming World Models

LOCI. Spatial Linear Memory for Streaming World Models. By Ji Xia and four others, from the Mohamed bin Zayed University of Artificial Intelligence and Pinscreen. Published on the thirtieth of September. Its code repository is public and real: fourteen source files.

Point a camera at a region, move away, and come back. A video world model should show what was there before. That takes two abilities: remembering past observations, and retrieving the right one for the current viewpoint. The two common forms of memory each give up one. A key-value cache keeps visual detail, but it grows with the length of the video. Recurrent memory stays compact, but it compresses history into a fixed-size state, after which individual past observations can no longer be reached directly. So one option remembers everything in detail at a cost that keeps climbing, and the other stays cheap but blurs the past together. For a model that must reproduce a specific place when the camera returns to it, neither is satisfying on its own.

LOCI is a hybrid spatial-memory architecture that declines to choose, and keeps both representations. The cache supplies detail. The recurrent memory supplies a compact running account of the scene. What makes the memory spatial is that the camera's geometry is built into how the recurrent part is read and written.

The transformer blocks are split in half. In one half, main attention keeps a key-value cache of past observations. In the other half, main attention is restricted to the current chunk, and is complemented by a recurrent linear-attention memory. That memory's reads and its writes are both conditioned on projective camera geometry, so the viewpoint shapes where things are stored and looked up, and also what is stored. The halves are connected. Readouts from the recurrent memory flow into the cache-backed blocks that follow, and give their queries accumulated context about the scene.

Tests use the public MIND memory benchmark and held-out recorded trajectories. LOCI reproduces revisited content more faithfully than representative world models, and more faithfully than a full-softmax model trained on the same recipe. With full history, it lowers peak memory by about thirty percent relative to full softmax at equal length.

A second regime matters for streaming. With a bounded bank of retained observations instead of the full history, LOCI streams long videos at constant memory, so the cost no longer grows with video length. Under that same budget it still remains more faithful than full softmax. That is the trade the problem described, a cache that grows against a state that forgets, and here the bounded version keeps memory flat without giving up faithfulness to the baseline it was compared with.

Memory for a world model, by this design, is not one mechanism but two cooperating ones. A cache keeps the detail, a camera-aware recurrent state keeps the scene, and the faithful revisit comes from letting the second inform the first.