News · 2026-07-11

LaMem-VLA gives robots a working memory so they stop forgetting what they just did

A new robotics model called LaMem-VLA targets a basic but stubborn flaw in today's robot brains: they mostly react to whatever the camera sees right now, and effectively forget what they just did. LaMem-VLA fixes this by giving the robot a working memory -- weaving compressed recollections of its recent and longer-term experience directly into the model's internal representation -- so it can handle multi-step tasks that require remembering an earlier state, like 'find the key, then go back and open the door.'

Key facts

• The flaw: vision-language-action models suffer 'temporal short-horizon bias' -- they predict the next action mostly from the current observation, forgetting task history.
• The fix: a four-stage memory pipeline (Curator, Seeker, Condenser, Weaver) that injects compact latent memory tokens into the model's input.
• Result: stronger performance on the SimplerEnv and LIBERO manipulation benchmarks, especially on long-horizon tasks.
• Primary source: the LaMem-VLA paper.

Start with what a vision-language-action model is, because that's the class of system this improves. These models take in what a robot sees (vision) and what it's told to do (language) and output what to do next (action) -- they're the increasingly common way to drive general-purpose robots. Most of them make a quiet simplifying assumption: decide the next action based mainly on the current frame. That works for reactive tasks -- pick up the cup you can see -- but it creates what the authors call 'temporal short-horizon bias.' The robot can see what's in front of it, but it loses track of where it is in a sequence. Ask it to do something that unfolds over several steps, where step four depends on remembering step one, and it flounders.

The obvious fixes each have a flaw, and understanding why LaMem-VLA is different means seeing them. You could feed the model more past frames -- but that bloats the input, slows everything down, and still buries the useful history in noise. Or you could bolt on an external memory bank the model queries -- but then memory is a separate 'hint' the model consults, not part of how it actually thinks. LaMem-VLA instead makes memory native to the model's internal representation, through a four-part pipeline the authors give evocative names. A Curator sorts past experience into two stores: short-term (recent state transitions, immediate tactical history) and long-term (broader task experience and recurring patterns). A Seeker uses what the robot currently sees and is trying to do to query those stores for relevant evidence. A Condenser compresses the retrieved material into a handful of compact 'latent memory tokens.' And a Weaver injects those tokens directly into the model's input sequence, right alongside the current image and instruction.

The analogy is the difference between a goldfish and a person doing a chore. The goldfish reacts to each moment fresh; the person carries a running sense of 'I've already done the first two steps, so now I do the third.' LaMem-VLA gives the robot that running sense -- but efficiently, by summarizing the relevant past into a few tokens the model reads as if they were part of the current scene, rather than replaying the whole history. On the SimplerEnv and LIBERO manipulation benchmarks it shows clear gains, concentrated exactly where you'd expect: the long-horizon tasks that demand remembering an earlier state.

Why it matters is that memory is one of the missing pieces between impressive robot demos and robots that can do real, extended work. A robot that only reacts to the present can pick and place; a robot that remembers what it has done can execute a recipe, tidy a room in the right order, or resume a task after an interruption. It moves robotics from reflexive -- see, act -- to reflective -- see, remember, act. The honest caveat is that these are simulation and standard-benchmark results, and the gap between LIBERO and a messy real kitchen is exactly where robotics claims usually shrink. It also connects to a broader thread in the field about what embodied systems retain and forget -- the same fragility that shows up when robots forget basic skills after new training. A better working memory is a real step; whether it survives contact with the physical world is the test that still has to be run.

Primary source, verified: read the paper → (arXiv 2607.07608)

Ground Truth. (2026, July 11). LaMem-VLA gives robots a working memory so they stop forgetting what they just did. Ground Truth. https://groundtruth.day/news/lamem-vla-gives-robots-a-working-memory.html

@misc{groundtruth:lamem-vla-gives-robots-a-working-memory,
  title  = {LaMem-VLA gives robots a working memory so they stop forgetting what they just did},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/lamem-vla-gives-robots-a-working-memory.html}
}

Topics: robotics · vision-language-action · memory · embodied-ai · research