OpenNovelty.org

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.0Download Report PDF
Agent MemoryLatent ReasoningLLM Agent
Abstract:

Agent memory shapes how Large Language Model (LLM)-powered agents, akin to the human brain, progressively refine themselves through environment interactions. Existing paradigms remain constrained: parametric memory forcibly adjusts model parameters, and retrieval-based memory externalizes experience into structured databases, yet neither captures the fluid interweaving of reasoning and memory that underlies human cognition. To address this gap, we propose MemGen, a dynamic generative memory framework that equips agents with a human-esque cognitive faculty. It consists of a \textit{memory trigger}, which monitors the agent’s reasoning state to decide explicit memory invocation, and a \textit{memory weaver}, which takes the agent's current state as stimulus to construct a latent token sequence as machine-native memory to enrich its reasoning. In this way, MemGen enables agents to recall and augment latent memory throughout reasoning, producing a tightly interwoven cycle of memory and cognition. Extensive experiments across eight benchmarks show that MemGen surpasses leading external memory systems such as ExpeL and AWM by up to 38.2238.22\\%, exceeds GRPO by up to 13.4413.44\\%, and exhibits strong cross-domain generalization ability. More importantly, we find that without explicit supervision, MemGen spontaneously evolves distinct human-like memory faculties, including planning memory, procedural memory, and working memory, suggesting an emergent trajectory toward more naturalistic forms of machine cognition.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces MemGen, a dynamic generative memory framework that represents memory as latent token sequences integrated directly into agent reasoning processes. According to the taxonomy, this work sits in the 'Generative Latent Memory Frameworks' leaf under 'Memory Architecture and Organization'. Notably, this leaf contains only the original paper itself with no sibling papers, suggesting this represents a relatively sparse and potentially novel research direction within the broader field of memory-augmented LLM agents.

The taxonomy reveals that MemGen's closest conceptual neighbors include 'Dynamic Agentic Memory Systems' (one paper) and 'Procedural Memory Learning' (one paper), both within the same parent branch. The broader field shows active development in memory retrieval mechanisms and evolutionary learning, but the generative latent approach appears distinct from these directions. The taxonomy's scope notes explicitly exclude external database memories and parametric weight updates from this category, positioning MemGen as exploring a middle ground between retrieval-based and parametric memory paradigms.

Among the 22 candidates examined across three contributions, no clearly refuting prior work was identified. The core MemGen framework examined 10 candidates with zero refutable matches, while the Memory Trigger/Weaver components examined 2 candidates with similar results. The emergent memory hierarchy contribution also found no overlapping prior work among 10 examined candidates. This limited search scope suggests that within the top-22 semantically similar papers, the specific combination of generative latent memory with dynamic triggering mechanisms appears not to have direct precedents.

Based on the available signals from this limited literature search, MemGen appears to occupy a relatively unexplored position within the memory-augmented agent landscape. The taxonomy structure shows this as a sparse leaf with no immediate siblings, and the contribution-level analysis found no refuting work among examined candidates. However, these findings reflect only the top-22 semantic matches and may not capture the full breadth of related work in adjacent research communities or alternative formulations of similar ideas.

Taxonomy

Core-task Taxonomy Papers
12
3
Claimed Contributions
22
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: generative latent memory for self-evolving language model agents. The field centers on equipping language model agents with memory systems that enable them to learn, adapt, and improve autonomously over time. The taxonomy reveals five major branches: Memory Architecture and Organization addresses how memory is structured and stored, ranging from hierarchical designs like Hierarchical Procedural Memory[9] to generative frameworks such as those in Generative Agents[2]; Memory Retrieval and Augmentation focuses on mechanisms for accessing and utilizing stored information effectively; Memory Learning and Evolution explores how agents refine their memory through experience, as seen in works like Evo-Memory[5]; Agent Application Domains examines deployment contexts from interactive simulations to specialized tasks; and Agent Development Frameworks and Tools provides infrastructure for building and evaluating these systems. These branches are deeply interconnected, as architectural choices directly influence retrieval strategies and learning dynamics. Recent work shows particularly active development in generative memory frameworks and evolutionary learning mechanisms. A central tension emerges between structured memory architectures that impose explicit organization versus more fluid generative approaches that synthesize memory representations on demand. MemGen[0] sits squarely within the generative latent memory frameworks, emphasizing self-evolution through latent representations that can be dynamically generated and refined. This positions it closely alongside Evo-Memory[5], which similarly focuses on memory evolution, though MemGen[0] appears to place stronger emphasis on the generative aspect of memory construction rather than purely evolutionary refinement. Compared to earlier foundational work like Generative Agents[2], MemGen[0] likely advances the self-evolving dimension, while works such as Agentic Memory[1] explore complementary perspectives on how agents manage and leverage their memory stores across diverse interaction scenarios.

Claimed Contributions

MemGen: Dynamic Generative Memory Framework

The authors introduce MemGen, a framework that enables LLM agents to dynamically generate and integrate latent memory during reasoning. Unlike parametric or retrieval-based approaches, MemGen interweaves memory and reasoning through continuous monitoring and generative reconstruction, creating a fluid cognitive cycle similar to human cognition.

10 retrieved papers
Memory Trigger and Memory Weaver Components

The framework comprises two synergistic components: a reinforcement learning-trained memory trigger that determines when to invoke memory based on the agent's cognitive state, and a memory weaver that generates machine-native latent memory sequences to enrich reasoning without modifying the core LLM parameters.

2 retrieved papers
Emergent Human-like Memory Hierarchy

The authors demonstrate that MemGen autonomously develops functionally specialized memory types resembling human cognitive faculties. These include planning memory for high-level task organization, procedural memory for task-specific skills, and working memory for maintaining context coherence, emerging without explicit architectural design for these distinctions.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

MemGen: Dynamic Generative Memory Framework

The authors introduce MemGen, a framework that enables LLM agents to dynamically generate and integrate latent memory during reasoning. Unlike parametric or retrieval-based approaches, MemGen interweaves memory and reasoning through continuous monitoring and generative reconstruction, creating a fluid cognitive cycle similar to human cognition.

Contribution

Memory Trigger and Memory Weaver Components

The framework comprises two synergistic components: a reinforcement learning-trained memory trigger that determines when to invoke memory based on the agent's cognitive state, and a memory weaver that generates machine-native latent memory sequences to enrich reasoning without modifying the core LLM parameters.

Contribution

Emergent Human-like Memory Hierarchy

The authors demonstrate that MemGen autonomously develops functionally specialized memory types resembling human cognitive faculties. These include planning memory for high-level task organization, procedural memory for task-specific skills, and working memory for maintaining context coherence, emerging without explicit architectural design for these distinctions.

Table of Contents