MIT DAAAM: Robots That Remember What They Saw, Where, and When

MIT’s DAAAM gives robots a real-time, long-term memory built from a live 4D scene graph. A smart batching trick runs the slow description model about 10 times faster, and the system raises accuracy on one spatial question-answering benchmark by 53.6%.
artificial-intelligence
Author

Kabui, Charles

Published

2026-06-23

Keywords

robot-memory, 4d-scene-graph, mit-daaam, embodied-ai, spatio-temporal-memory

Download as PDF

Researchers at MIT’s SPARK Lab released DAAAM (Describe Anything, Anywhere, at Any Moment), a system that gives robots a real-time, long-term memory of what they saw, where, and when. As a robot moves, DAAAM turns the feed from a depth-sensing camera, which captures color and distance, into a 4D scene graph: a constantly updated database where every object and place gets a written description, a 3D position, and a timestamp. The model that writes those descriptions is slow, so DAAAM picks the clearest frames of each object and runs them through in batches on a separate thread. That speeds the step up roughly tenfold, making real-time operation possible while the robot tracks objects at the camera’s full 10 frames per second.

A robot that forgets cannot handle long jobs. With DAAAM, a separate AI agent searches that memory in plain language, so a warehouse or hospital robot can answer “where did you last see the red tool?” or “how long has that door been open?” Against the best existing methods, DAAAM answered questions more accurately, raising accuracy on one benchmark by 53.6%, and completed step-by-step navigation tasks correctly about 28% more often. The code and data are open-source.

AI’s text agents spent the year racing to build better memory. DAAAM moves that race into the physical world, where the limit is shifting from how well a robot reads a single frame to how well it remembers a place over time. It still trips on unusual objects and fast-moving drones.

Read More: RynnBrain is the other half of this push, one open model that lets a robot perceive, reason, and act.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {MIT {DAAAM:} {Robots} {That} {Remember} {What} {They} {Saw,}
    {Where,} and {When}},
  date = {2026-06-23},
  url = {https://toknow.ai/posts/mit-daaam-robot-real-time-memory-4d-scene-graph/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “MIT DAAAM: Robots That Remember What They Saw, Where, and When.” https://toknow.ai/posts/mit-daaam-robot-real-time-memory-4d-scene-graph/.