MemAct / StreaMemAct

Paper · Project · Code

Yuxi Wei, Wei Huang, Qirui Chen, Lu Hou, Xiaojuan Qi.

This repository hosts the prepared StreaMemAct data suite accompanying MemAct: Streaming Spatial Understanding with Memory and Active Exploration. The repository identifier is vfishc/memact. The repository is public; data upload is in progress.

Contents

Component	Count / scope	Directory
Current QA	302,661 answer turns	data/train/current_memory
Memory QA	402,662 answer turns	data/train/current_memory
Memory supervision	519,948 writes; 556,635 retrieval calls, additional to QA counts	Within Current/Memory streams
Action SFT	294,677 items; 534,752 camera transitions	data/train/action
RL	14,001 tasks and branch observations	data/rl
Main evaluation	20,000 turns, including graphs for 1,245 Action questions	data/evaluation
Timing	2,000 separate questions	data/evaluation
Simulation media and GLBs	Rendered videos and associated scene assets	media/simulation
Rendering dependencies	Scene inputs, textures/HDRIs and configurations	assets/rl
Frozen Action render inputs	1,245 question configurations	generation/frozen_action
Metadata	Frames, timestamps, camera/object records and evidence references	metadata/episodes
Reconstruction	Upstream download/reconstruction mappings	reconstruction

The prepared manifest records 206 archives totaling 589,652,365,064 bytes, plus supporting metadata. Archives retain relative media references and checksums. See release_manifest.json, release_statistics.json and RELEASE_GUIDE.txt after upload for extraction and verification.

Usage and provenance

Use the code repository's generation, reconstruction, training and evaluation entry points. Retrieve original ScanNet, ScanNet++ and ARKitScenes recordings from their providers; this release supplies reconstruction mappings rather than redistributing original recordings. A video does not replace its frame/time mapping. Evaluator-only answers and geometry must not be included in model inputs. Weights are not included.

The prepared training selection contains 1M QA turns; memory writes and retrieval calls are counted separately. It replaces selected recording-ID overlap with the frozen evaluation bank. Strict physical-space disjointness is not claimed because distinct real recordings can share a physical space. Release counts are separate from paper-era intermediate composition.

Technical validation covers annotation/media references, branch coverage, reconstruction samples and runtime tests. It is not a rerun of the paper's benchmark scores. The supplied native inference backend does not reproduce the persistent cross-call compact KV cache or paper latency.

Licenses

See LICENSE.md. Original MemAct annotation contributions use CC BY 4.0, and original code uses Apache-2.0. Upstream assets and derived content retain their source terms. The mixed repository is not licensed wholesale under CC BY 4.0; applicable asset-specific source terms remain controlling.