Memory-Augmented World Model for Green AI

This project proposes a Memory-Augmented World Model to improve long-horizon consistency in agentic decision-making and world prediction. Rather than relying only on scaling large parametric backbones, the team introduces an explicit non-parametric memory bank that stores curated common sense and transition knowledge. During training and rollout, the model performs batched ANN retrieval to condition predictions on stable memory “anchors,” reducing error accumulation and drift over long trajectories. The team will evaluate on ALFWorld, WebShop, WebArena, and additional agentic benchmarks, reporting task success, step efficiency, and robustness under long-horizon rollouts, together with visual consistency proxies (identity/attribute stability, temporal coherence). The project will also track practical training and deployment metrics relevant to EuroHPC usage, including time-to-target performance, total node/GPU hours, and distributed scaling behaviour under PyTorch DDP/FSDP. Retrieval overhead is managed via index sharding/caching and systematic ablations over memory size, retrieval frequency, and filtering strategies. Deliverables include reproducible training/evaluation scripts, technical reports, and publication-ready results.