Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents

2026-03-15 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

The "Experience Compression Spectrum" is a new framework that unifies agent memory, skill discovery, and rule learning for LLM agents, addressing the critical bottleneck of managing accumulated experience in long-horizon deployments. It positions memory, skills, and rules along a single axis of increasing compression, ranging from 5–20x for episodic memory (Level 1) to 50–500x for procedural skills (Level 2), and over 1,000x for declarative rules (Level 3). This compression directly reduces context consumption, retrieval latency, and compute overhead. A citation analysis of 1,136 references across 22 papers revealed a cross-community citation rate below 1% between memory and skill research. The framework maps over 20 existing systems, showing that each operates at a fixed compression level, highlighting a "missing diagonal" where no system supports adaptive cross-level compression. Higher compression consistently yields better downstream performance, but current systems lack automated rule extraction and adaptive level selection.

Key takeaway

For AI Engineers designing scalable LLM agents, recognize that efficient experience management requires moving beyond single-level memory or skill systems. Your designs should aim for adaptive, full-spectrum compression, allowing agents to dynamically promote and demote knowledge across episodic memories, procedural skills, and declarative rules to optimize for context, retrieval, and compute. Prioritize building systems that can automatically extract and manage L3 rules, as this is a significant gap in current agent architectures.

Key insights

Experience compression unifies LLM agent memory, skills, and rules, crucial for scalable, long-horizon deployments.

Principles

• Higher compression reduces context, latency, and compute.
• Transferability increases with compression, specificity decreases.
• Acquisition cost vs. maintenance cost trade-off.

Method

The Experience Compression Spectrum formalizes four levels: Raw Trace (L0), Episodic Memory (L1, ~5-20x), Procedural Skill (L2, ~50-500x), and Declarative Rule (L3, ~1000x+), each with distinct formats and reusability.

In practice

• Distill traces into abstract skills for performance gains.
• Consider multi-level knowledge maintenance for scale.
• Automate rule extraction for L3 knowledge.

Topics

• Experience Compression Spectrum
• LLM Agents
• Agent Memory Systems
• Agent Skill Discovery
• Declarative Rules

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

Related on AIssential

• Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents — The Experience Compression Spectrum unifies LLM agent memory, skills, and rules by their compression levels.
• 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.
• Agents Fail to Reject Stale Memories — LLM efficiency, steerability, and agent memory coherence face fundamental limitations requiring novel architectural solutions.
• AEL: Agent Evolving Learning for Open-Ended Environments — AEL enables LLM agents to learn from experience by dynamically selecting memory retrieval policies and integrating reflective causal insights.
• 🥇Top AI Papers of the Week — AI advancements focus on novel computing, efficient memory, robust agents, and specialized models.
• 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.

Open in AIssential →

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