# LTM gives robots a compact memory of past actions

> Researchers introduce a way for vision-language-action robots to retain past interactions as compact transitions, reporting higher success on memory tasks.

Oossa · 2026-10-16 · https://oossa.com/en/ltm-gives-robots-a-compact-memory-of-past-actions

Researchers introduce Latent Transition Memory (LTM), a method that helps robots remember interactions that are no longer visible. Instead of storing full observations, it records compact transitions between consecutive observations and uses them with the π₀.₅ vision-language-action model.

In their reported tests, LTM-π reached 80.5% success on LIBERO-Mem and 58.54% on RoboMME, outperforming baselines that used neither external reasoning models nor memory-specific annotations. The paper also reports real-robot counting success rising from 11.25% to 82.50%; the abstract does not give details about the robot or test setup.

## The facts

- The method stores transitions between consecutive observations and uses causal KV caching for incremental inference.
- The paper reports 84.00% success with a 200-token memory budget, compared with 87.50% without compression.

## Why it matters

For robotics researchers, LTM offers a tested way to fit interaction history into a small memory budget, though the abstract alone does not establish how well it transfers to other robots or tasks.

## Sources & references

1. [LTM: Latent Transition Based In-Context Memory for Vision-Language-Action Models](https://arxiv.org/abs/2610.21362) – arXiv cs.RO, 2026-10-16

Last updated: 2026-10-16