Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents

Hugging Face Daily Papers Papers

Summary

Introduces Belief Entropy and Metacognitive Memory Policy Optimization (MMPO) to improve memory quality in long-horizon LLM agents, outperforming existing methods and maintaining performance over long contexts.

Memory-augmented LLM agents tackle complex long-horizon tasks by recursively summarizing interaction trajectories into compact memory. However, existing approaches typically train these memory policies using outcome-based reinforcement learning, failing to localize where intermediate memory quality degrades. As interactions unfold, ambiguous recursive summaries progressively discard task-relevant information and introduce semantic noise. This exacerbates belief deviation, obscuring the agent's estimate of the latent task state and ultimately derailing long-horizon reasoning. We therefore argue that memory optimization should focus not merely on trajectory-level success, but on the clarity of the belief induced by intermediate summaries. To this end, we introduce Belief Entropy, a self-supervised proxy that probes how uncertain the model remains about the latent task state given its current memory. Based on this proxy, we propose Metacognitive Memory Policy Optimization (MMPO). Instead of relying only on sparse outcome-based signals, MMPO provides fine-grained, memory-specific supervision via explicitly penalizing summaries that induce high epistemic uncertainty. Experiments show that MMPO consistently outperforms existing methods on diverse long-horizon tasks, maintaining 97.1% performance even when scaled to 1.75M-token contexts.
Original Article
View Cached Full Text

Cached at: 06/05/26, 06:07 AM

Paper page - Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents

Source: https://huggingface.co/papers/2605.30159 Published on May 28

·

Submitted byhttps://huggingface.co/ziyan2003

liuon Jun 5

Abstract

Memory-augmented language models struggle with long-horizon tasks due to information loss in recursive summaries, but a new method using belief entropy and metacognitive policy optimization improves performance by focusing on memory quality rather than just outcome success.

Memory-augmented LLM agentstackle complexlong-horizon tasksby recursively summarizing interaction trajectories into compact memory. However, existing approaches typically train these memory policies using outcome-basedreinforcement learning, failing to localize where intermediate memory quality degrades. As interactions unfold, ambiguous recursive summaries progressively discard task-relevant information and introduce semantic noise. This exacerbatesbelief deviation, obscuring the agent’s estimate of thelatent task stateand ultimately derailing long-horizon reasoning. We therefore argue that memory optimization should focus not merely on trajectory-level success, but on the clarity of the belief induced by intermediate summaries. To this end, we introduceBelief Entropy, a self-supervised proxy that probes how uncertain the model remains about thelatent task stategiven its current memory. Based on this proxy, we proposeMetacognitive Memory Policy Optimization(MMPO). Instead of relying only on sparse outcome-based signals, MMPO provides fine-grained, memory-specific supervision via explicitly penalizing summaries that induce highepistemic uncertainty. Experiments show that MMPO consistently outperforms existing methods on diverselong-horizon tasks, maintaining 97.1% performance even when scaled to 1.75M-token contexts.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.30159

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.30159 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.30159 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.30159 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Belief Memory: Agent Memory Under Partial Observability

arXiv cs.AI

This paper introduces BeliefMem, a novel memory paradigm for LLM agents that stores multiple candidate conclusions with probabilities to handle partial observability and reduce self-reinforcing errors. Empirical evaluations show it outperforms deterministic baselines on LoCoMo and ALFWorld benchmarks.

MemGym: a Long-Horizon Memory Environment for LLM Agents

arXiv cs.CL

MemGym is a benchmark for evaluating memory formation in LLM agents over long-horizon tasks, unifying existing agent gyms and synthetic pipelines with memory-isolated scores. It spans tool-use dialogue, multi-turn search, coding, and computer use, and includes a lightweight reward model (MemRM) for efficient evaluation.