HeLM: Hierarchical Explicit Language Memory for Non-Markovian Manipulation

HeLM. Hierarchical Explicit Language Memory for Non-Markovian Manipulation. By Gyeonghun Kim and three others, from Seoul National University. Published on the second of November.

Vision-Language-Action models can now perform fine manipulation, but tasks that need working memory remain a key challenge. Most of these architectures operate under a Markovian assumption, mapping only the current observation to an action. In a non-Markovian environment, where the right action depends on what happened earlier, that assumption leads to failure.

HeLM, short for Hierarchical Explicit Language Memory, takes its inspiration from the human inner monologue. It is a hierarchical VLA framework that keeps task-relevant history in explicit language, written down rather than held implicitly.

Memory is maintained through two pieces working together: factorized event detection, and memory updates that are triggered by those events. The record changes when something happens, not at every step. Because the memory is language, what the policy carries forward is a compact account of the events that matter to the task, and that account can be kept across the boundary between one episode and the next.

Evaluation covers ten non-Markovian real-world tasks with fifty-six variants. Five of those tasks are newly formulated inter-episode tasks, which require context to persist across task boundaries. HeLM improves intra-episode success by twelve point eight percentage points over the strongest baseline, and reaches seventy-eight percent subtask success on compositional generalization tasks.

Against the keyframe baseline at long horizons, HeLM uses up to four point five times less memory and runs two point five times faster. Writing the past down in a few words, only when events call for it, turns out to be both the lighter record to keep and the one that lets the robot act on history.