Latent Dynamics · 2026-10-09
MemoWM: How World Models Change What Agents Need to Remember
Bingfan Zeng, Zhisheng Chen, Chenbo Sang, Zhengwei Xie, Jinpeng Wang, Xiangchen Guan, Rui Qian, Zheng Lu, Jingwei Song
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
MemoWM is a long-term episodic memory framework that uses a shared, action-conditioned world model as a predictive prior to compress stored experience. It allocates what to retain per discrete coordinate by balancing expected task impact of reconstruction errors against storage cost, then reconstructs omitted content in a closed loop without transmitting retention masks. Across five benchmarks, it reports 42.42% average answer accuracy and 53.9% lower experience-specific storage vs. MIRIX, while releasing code.
Why it matters
The paper frames memory as a division of labor between reusable world-model knowledge and per-experience episodic storage. By turning world-model predictions into a prior that both codes retained information and reconstructs omitted coordinates, MemoWM aims to reduce how much each interaction must be stored while maintaining task performance. This enables a task-aware memory allocation rule that does not require future queries and quantifies a trade-off between world-model capacity and recurring memory cost.
Method
- Discretize each observation into a query-independent state and use an action-conditioned world model to predict per-coordinate distributions from a reconstructed history.
- Estimate omission risk per coordinate via offline single-coordinate counterfactual readouts measuring how prediction errors change answer loss, then retain coordinates whose expected task impact exceeds predicted storage cost (threshold λ).
- Reconstruct omitted coordinates from the shared modal prediction and feed the reconstructed states back into the next prediction, so writer and decoder stay synchronized without transmitting a per-state retention mask.
Limitation
The MemEye experiment assesses memory on synthetic data constructed from the public benchmark’s screen/task structure, and “does not establish zero-shot generalization to unseen interface styles.”
Abstract (from arXiv)
Long-term agents face growing storage demands as they accumulate experience. World models capture reusable regularities that can reduce the information stored for each experience. We formulate the problem of memory allocation conditioned on a world model and introduce MemoWM, a framework that uses shared predictions to compress retained information and reconstruct omitted content. Its task-aware allocation rule balances the expected impact of reconstruction errors against storage cost, retaining information with downstream value beyond the predictive prior. Across five long-term agent-memory benchmarks, MemoWM achieves 42.42\% average answer accuracy, exceeding the strongest baseline by 2.62 percentage points, while reducing average experience-specific storage by 53.9\% relative to MIRIX, the most storage-efficient baseline. Further analysis shows that stronger world models reduce per-experience storage at comparable task quality. Accounting for model parameters reveals a trade-off between shared model capacity and recurring storage costs, with the capacity that minimizes total storage increasing as more interactions are retained. Our code is available at https://github.com/Feld-maxiu/MemoWM.
Related papers
- Heavy-Tailed Memory Traces in Long-Horizon Language Agents
- Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
- EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
- Self-Improving Agents with World Model Predictive Paths
- What Should World Models Forget? Stratified Retention for Continual Adaptation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-09 UTC
MemoWM: How World Models Change What Agents Need to Remember
Preprint posted: 2026-10-07 · Original record ↗
- arxiv
- 2610.10778
Explicit source versions (1)
- v1 · 2026-10-07
- Preprint available
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Long-term agents face growing storage demands as they accumulate experience. World models capture reusable regularities that can reduce the information stored for each experience. We formulate the problem of memory allocation conditioned on a world model and introduce MemoWM, a framework that uses shared predictions to compress retained information and reconstruct omitted content. Its task-aware allocation rule balances the expected impact of reconstruction errors against storage cost, retaining information with downstream value beyond the predictive prior. Across five long-term agent-memory benchmarks, MemoWM achieves 42.42\% average answer accuracy, exceeding the strongest baseline by 2.62 percentage points, while reducing average experience-specific storage by 53.9\% relative to MIRIX, the most storage-efficient baseline. Further analysis shows that stronger world models reduce per-experience storage at comparable task quality. Accounting for model parameters reveals a trade-off between shared model capacity and recurring storage costs, with the capacity that minimizes total storage increasing as more interactions are retained. Our code is available at https://github.com/Feld-maxiu/MemoWM.