菜单
Meta-Cognitive Memory Policy Optimization (MMPO) for long-horizon LLM agents using Belief Entropy as self-supervised proxy.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/hiyenwong/ai_collection --skill mmpo-metacognitive-memory-policy命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | mmpo-metacognitive-memory-policy |
| description | Meta-Cognitive Memory Policy Optimization (MMPO) for long-horizon LLM agents using Belief Entropy as self-supervised proxy. |
| arxiv_id | 2605.30159 |
| authors | Ziyan Liu, Zhezheng Hao, Yeqiu Chen, Hong Wang |
| published | 2026-05-29T00:00:00.000Z |
| categories | ["Artificial Intelligence","LLM Agents","Memory Optimization","Metacognitive"] |
| tags | ["LLM","agents","memory","reinforcement-learning","belief-entropy","long-horizon","metacognitive"] |
| activation_keywords | ["memory policy optimization","belief entropy","long-horizon LLM agent","metacognitive","MMPO","memory-augmented agent"] |
| related_skills | ["agent-memory-framework","agent-memory-management","indexed-memory"] |
Meta-Cognitive Memory Policy Optimization (MMPO) is a novel approach for training memory-augmented LLM agents on long-horizon tasks. It addresses the critical problem of belief deviation in recursive memory summarization by introducing Belief Entropy — a self-supervised proxy that measures epistemic uncertainty about the latent task state.
Key Innovation: Instead of relying solely on sparse outcome-based reinforcement learning signals, MMPO provides fine-grained, memory-specific supervision by explicitly penalizing summaries that induce high epistemic uncertainty.
arXiv: 2605.30159
Memory-augmented LLM agents tackle complex long-horizon tasks by recursively summarizing interaction trajectories into compact memory. However:
Belief Entropy is a self-supervised proxy that probes how uncertain the model remains about the latent task state given its current memory:
Metacognitive Memory Policy Optimization consists of:
Memory Summarization
Belief Entropy Estimation
Memory-Specific Supervision
Policy Optimization
# Conceptual implementation (not from paper)
def belief_entropy(memory, task_context):
"""Estimate epistemic uncertainty about task state."""
belief_distribution = model.predict_task_state(memory)
entropy = -sum(p * log(p) for p in belief_distribution)
return entropy
def mmpo_loss(outcome_reward, memory, belief_entropy, lambda_penalty):
"""Combined loss function."""
rl_loss = -outcome_reward # Sparse RL signal
entropy_penalty = lambda_penalty * belief_entropy(memory)
return rl_loss + entropy_penalty
Performance Highlights:
Comparison with Existing Methods:
| Method | Context Scaling | Performance | Memory Quality |
|---|---|---|---|
| Outcome-based RL | ~500K tokens | Degrades | Poor localization |
| MMPO | 1.75M tokens | 97.1% retention | High clarity |
Task-Relevant Information Preservation
Semantic Noise Reduction
Self-Supervised Supervision
Belief Clarity Focus
Over-reliance on Belief Entropy
Incorrect Entropy Estimation
Ignoring Task Structure
Scaling Challenges
MMPO introduces Belief Entropy as a self-supervised proxy for memory policy optimization in long-horizon LLM agents. By penalizing high-uncertainty memory summaries, it provides fine-grained supervision that outcome-based RL lacks. This enables agents to maintain task-relevant information clarity over extreme context scales (1.75M tokens), achieving 97.1% performance retention.
Core Principle: Optimize memory for belief clarity, not just task outcomes.