Memp: Exploring Agent Procedural Memory

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

$Mem^{p}$ is a task-agnostic framework designed to equip Large Language Model (LLM)-based agents with learnable, updatable, and lifelong procedural memory. It addresses the brittleness of existing procedural memory in agents by distilling past trajectories into both fine-grained, step-by-step instructions and higher-level, script-like abstractions. The framework systematically explores strategies for building, retrieving, and updating this memory, including dynamic regimens for continuous correction and deprecation. Empirical evaluations on the TravelPlanner and ALFWorld benchmarks demonstrate that as the memory repository is refined, agents achieve consistently higher success rates and greater efficiency on analogous tasks. Notably, procedural memory built from stronger models like GPT-4o can be transferred to weaker models such as Qwen2.5-14B-Instruct, yielding substantial performance gains, including a 5% increase in task completion rate and a 1.6-step reduction on TravelPlanner.

Key takeaway

For Research Scientists developing LLM-based agents, integrating a dynamic procedural memory system like $Mem^{p}$ is crucial for improving agent robustness and efficiency. You should focus on implementing memory construction that combines abstract scripts with concrete trajectories, employ semantic-aware retrieval, and prioritize reflection-based update mechanisms to ensure continuous learning and adaptation, especially for long-horizon and complex tasks.

Key insights

Procedural memory, dynamically built and updated, significantly enhances LLM agent performance and efficiency across diverse tasks.

Principles

• Procedural memory improves task accuracy and reduces execution steps.
• Abstracted scripts generalize better than raw trajectories.
• Procedural memory is transferable between models of varying strength.

Method

$Mem^{p}$ distills agent trajectories into fine-grained instructions and high-level scripts. It employs strategies for memory construction (trajectories, scripts, combined), retrieval (query, AveFact), and dynamic updating (vanilla, validation, adjustment/reflection).

In practice

• Combine abstract guidelines with concrete trajectories for optimal memory.
• Use query-based or AveFact methods for precise memory retrieval.
• Implement reflection-based updates for continuous memory refinement.

Topics

• Procedural Memory
• LLM Agents
• Memory Management
• Trajectory Distillation
• Memory Transfer

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• MemRefine: LLM-Guided Compression for Long-Term Agent Memory — LLM-guided memory compression, prioritizing factual content over surface similarity, effectively manages agent memory within fixed budgets.
• MemCollab: Cross-Agent Memory Collaboration via Contrastive Trajectory Distillation — MemCollab enables LLM agents to share memory by distilling agent-agnostic reasoning constraints from diverse reasoning trajectories.
• AI Agents of the Week: Papers You Should Know About — Advancements in AI agents focus on improving memory, planning, collaboration, and reliability through novel architectures and practical tools.
• Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents — Agent memory should manage execution state hierarchically, not just retrieve semantically similar facts.
• Multi-Agent Transactive Memory — MATM allows LLM agents to share procedural trajectories, improving task performance and reducing steps without coordination or joint training.
• Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution — ReMe dynamically manages LLM agent procedural memory, enabling smaller models to surpass larger, memoryless counterparts through experience-driven evolution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.