Nous: A Predictive World Model for Long-Term Agent Memory

arXiv 2026 49.4 method

TLDR

Nous uses a predictive world model for agent memory, storing belief updates instead of facts, and achieves strong results on a conversational benchmark.

Reasoning

The paper introduces a novel memory architecture based on predictive world models and Bayesian surprise, with empirical evaluation on the LoCoMo benchmark. Strengths include a principled approach to forgetting and identity resolution; weaknesses include limited scope to conversational memory and reproducibility concerns in baseline comparisons.

Read-first score

Read-first score 49.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 10.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

94.6

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94582622

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation,metric,result

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=artifact

Topical relevance 29%

14.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 358.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Predictive world model based on knowledge as prediction, not storage
Categorical probability distributions (dimensions) per entity-attribute pair
Scoring observations by information-theoretic surprise S = -log2 P(obs | D) and updating via closed-form Bayesian posterior
Storing delta (shift from prior to posterior belief) rather than the fact itself
Forgetting via entropy decay toward uniform distribution
Identity resolution through mutual information between entity dimension sets

Methodology

Nous maintains a predictive world model as a collection of categorical probability distributions (dimensions) for each entity-attribute pair observed in conversation. Incoming observations are scored by surprise S = -log2 P(obs | D) and the distribution is updated via a closed-form Bayesian posterior, storing only the delta (shift from prior to posterior). Forgetting emerges from entropy decay toward uniform, and identity resolution uses mutual information between entity dimension sets. The model is evaluated on the LoCoMo benchmark (10 conversations, 1,540 questions) using GPT-4o-mini as backbone, with F1 metrics for single-hop, multi-hop, temporal, and open-domain categories, compared against A-MEM and BeliefMem.

Key Results

Nous achieves F1 scores of 63.50 (single-hop), 55.32 (multi-hop), 58.57 (temporal), and 62.50 (open-domain) on the LoCoMo benchmark. It shows substantial gains over A-MEM in three of four categories and exceeds BeliefMem's self-reported numbers on all four categories, though comparisons are limited by reproducibility and pipeline differences.

Limitations

Reproducibility issue: independent citations of A-MEM's results disagree on category assignment, which the authors discuss openly rather than resolve unilaterally
Uncontrolled differences between evaluation pipelines prevent the comparison with BeliefMem from being fully controlled

Resources

Paper