memory-guided iterative generation

This paper addresses the lack of formal modeling of memory mechanisms in LLM-based agents. Methodologically, it introduces the first unified, dynamic memory analysis framework, decomposing memory into three orthogonal representations—parametric, structured, and unstructured—and six atomic operations: consolidation, update, indexing, forgetting, retrieval, and compression. Through a systematic literature review and operation–theme mapping, the framework unifies research strands including long-term memory, long-context modeling, parameter editing, and retrieval-augmented generation (RAG). Its contributions are threefold: (1) construction of a comprehensive knowledge graph spanning all memory dimensions, integrating over 100 methods, benchmarks, and tools; (2) precise functional characterization and coordination pathways for each atomic operation; and (3) the first formal memory modeling foundation for LLM agents—enabling interpretable, scalable memory system design with both theoretical grounding and practical guidance.

This study addresses the critical gap in governance of persistent state—such as memory, credentials, and commitments—in long-running large language model agents, particularly concerning recoverability, auditability, and controlled deprecation. Through a systematic review of 435 publications, the work introduces the first comprehensive model of agent persistent state, integrating multidimensional elements including task logs, credentials, and commitments. Building on this foundation, it proposes AOEP-v0, an evaluation protocol grounded in six dimensions: authoritativeness, scope, mutability, provenance, recoverability, and operability. Distinct from prior approaches that prioritize response quality alone, AOEP-v0 explicitly centers on obligations surrounding state modification and recovery, establishing the first cross-domain governance benchmark for the reliability and controllability of always-on intelligent agents.

This work identifies distributional drift and performance degradation induced by synthetic-data-based cyclic self-training in continual learning, specifically within the generative experience replay (GER) paradigm, where model reliability and latent-space alignment are systematically compromised. Methodologically, we provide the first statistical proof that synthetic data introduce substantial bias and variance, undermining the consistency of maximum likelihood estimation; we further uncover an implicit collapse phenomenon in mainstream generative models (GANs/VAEs) during iterative self-training. Through rigorous statistical modeling, quantitative measurement of latent-space alignment, and multi-round GER experiments, we empirically demonstrate that all evaluated methods suffer over 60% degradation in latent-space alignment and a 2.3× increase in reconstruction error after just 3–5 self-training cycles. These findings establish theoretical foundations and empirical warnings regarding the safety and stability of GER in continual learning systems.

To address the dual challenges of skill reusability and catastrophic forgetting in continual learning, this paper proposes a “what–how” dual-system recursive neural network architecture that decouples task selection (what) from execution policy (how), enabling flexible skill composition and incremental updates. Methodologically, it introduces the first integration of online unsupervised task-structure inference with context-driven dynamic assembly of low-rank RNN components: a probabilistic generative model captures time-varying task structure, which then modulates the real-time composition of functional low-rank RNN modules. Experiments demonstrate that the framework significantly mitigates forgetting in multi-task cognitive settings, enables strong forward and backward transfer, and achieves rapid generalization to unseen tasks. This work establishes an interpretable, scalable paradigm for continual learning endowed with cognitive flexibility.

This work addresses catastrophic forgetting in continual learning by proposing a "learning-in-dreams" mechanism inspired by human dreaming. Without relying on replaying real samples, the method autonomously generates semantically novel yet knowledge-consistent structured synthetic data through a frozen diffusion model guided by classifier-driven soft prompt optimization. This approach actively reorganizes and reconstructs the representational space to balance model plasticity and stability, enabling forward-looking self-training and positive forward transfer. Experiments on Mini-ImageNet, FG-ImageNet, and ImageNet-R demonstrate that the proposed method significantly outperforms strong baselines, substantially enhancing the model’s adaptability across sequential tasks.

Current AI architectures lack human-like cognitive properties—specifically, working memory with sustained representational continuity, dynamic mental imagery generation, and coherent mental state evolution across time. Method: We propose a biologically inspired cognitive architecture integrating Global Workspace Theory (GWT) with hierarchical neural networks. Crucially, we embed a bimodal persistent activity mechanism—comprising sustained neural firing and activity-dependent synaptic plasticity—directly into the global workspace, enabling dynamic maintenance, iterative updating, and cross-state continuous evolution of short-term representations. Mental imagery and gradual working memory evolution are modeled via continuous-time neural dynamics and iterative state updates. Contribution/Results: The model successfully simulates human-like mental imagery generation, progressive working memory transformation, and reasoning transitions. Experiments demonstrate its capacity to support mental continuity modeling and the emergence of general intelligence, offering a cognitively interpretable architectural foundation for Artificial General Intelligence.

This work addresses the limitations of existing AI agent workflows, which rely on implicit dialogue states and struggle to ensure stability of intermediate artifacts, isolate irrelevant updates, and propagate changes precisely. To overcome these challenges, the paper proposes modeling AI-native workflows as directed acyclic graphs (DAGs) and introduces the concept of execution lineage. By leveraging explicit dependency tracking, identity-based identification of intermediate artifacts, and an identity-aware replay mechanism, this approach achieves deterministic computation graphs in AI agents for the first time. The method guarantees precise change propagation, zero contamination across unrelated branches, and preserves both upstream stability and cross-artifact consistency. Evaluated on a policy memo updating task, DAG-based replay attains 100% fidelity in final outputs, substantially outperforming iterative baseline approaches.

Existing generative world models exhibit poor performance in backtracking simulations—a limitation often mistakenly attributed to model capacity but actually stemming from deficiencies in runtime state management. This work proposes a session-centric runtime architecture that explicitly distinguishes between recomputable and non-recomputable states, introducing the concept of the minimal non-recomputable state (PCS) and enabling its efficient snapshotting and restoration. By capturing observations, random number generator states, memory banks, and sliding KV contexts, and employing a relevance-based rather than recency-based memory eviction policy, the system achieves state recovery in just 0.012 milliseconds—orders of magnitude faster than the 1.85 seconds required per generation step. The approach supports up to 1,024 concurrent sessions while ensuring exact backtracking fidelity and significantly enhancing scalability.

Large language models struggle to convert reasoning trajectories into reusable knowledge due to their stateless nature, hindering continuous self-improvement. This work proposes an unsupervised online learning framework that distills raw reasoning traces into lightweight, modular latent memories using self-generated signals—such as majority voting—as implicit rewards. A minimal soft-prompt mechanism enables efficient storage and retrieval of these memories. Combined with a few gradient-update steps for fine-tuning, the approach substantially outperforms both zero-shot and trajectory-augmented in-context learning baselines across multiple mathematical reasoning benchmarks, achieving performance on par with full-parameter fine-tuning and offline training while demonstrating strong cross-dataset transferability.

Current text-to-image generation models struggle with implicit visual constraints, relational reasoning, and prompts requiring external knowledge, while also failing to effectively leverage historical generation experience. This work proposes MemoGen—a training-free, test-time self-evolution framework that introduces, for the first time, a reusable experience memory mechanism into image generation. By employing an agent evolution layer, MemoGen establishes a closed-loop pipeline that integrates task comprehension, external evidence retrieval, constraint formulation, result evaluation, and memory storage, enabling continuous refinement of generation strategies without updating model parameters. Built upon the Qwen-Image backbone, MemoGen surpasses strong baselines such as Nano Banana Pro and GPT-Image-1 on knowledge- and reasoning-intensive benchmarks like WISE and Mind-Bench after only two rounds of evolution.