Mimir: A Neuro-Symbolic Memory System with Dynamic Grounding for Embodied Agents in Interactive Environments
Mimir is a neuro-symbolic memory system for long-horizon embodied tasks under partial observability. It separates world memory, which stores object locations, states, and perceptual evidence, from task memory, which tracks the ordered goal agenda, progress, hand state, failures, and execution constraints. Before each action, a grounding module connects the active goal to recalled world candidates, fills missing source locations, and attaches evidence for planning and execution. The paper reports maximum gains of 42.5% on EB-ALFRED and 23.0% on EB-Habitat, an 8.5% overall success-rate improvement over prior systems under the same backbone, and 86.0% success on the EB-Habitat long-horizon subset.
Why it's worth reading
Mimir targets a concrete bottleneck in long-horizon embodied agents: memories are often disconnected from the active goal. Its explicit task/world split and evidence-based grounding are paired with reported gains across EB-ALFRED and EB-Habitat.