Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents
Summary: arXiv:2604.15877v1 Announce Type: new
Abstract: As LLM agents scale to long-horizon, multi-session deployments, efficiently managing accumulated experience becomes a critical bottleneck. Agent memory systems and agent skill discovery both address this challenge — extracting reusable knowledge from interaction traces — yet a citation analysis of 1,136 references across 22 primary papers reveals a cross-community citation rate below 1%. We propose the Experience Compression Spectrum, a unifying framework that positions memory, skills, and rules as points along a single axis of increasing compression (5–20× for episodic memory, 50–500× for procedural skills, 1,000×+ for declarative rules), directly reducing context consumption, retrieval latency, and compute overhead.
Introduction
The growth of large language models (LLMs) has introduced new challenges in the field of artificial intelligence. As these agents become increasingly complex and deployed over extended periods, efficient management of their accumulated experiences is paramount. The current landscape reveals a fragmented approach, with various solutions being developed independently across different communities.
The Experience Compression Spectrum
In light of these challenges, we introduce the Experience Compression Spectrum, which serves as a comprehensive framework to understand the relationship between memory, skills, and rules in LLM agents. This spectrum categorizes these elements based on their compression capabilities:
- Episodic Memory: 5-20× compression
- Procedural Skills: 50-500× compression
- Declarative Rules: 1,000×+ compression
This framework not only highlights the varying degrees of compression but also emphasizes how they can contribute to reducing context consumption, retrieval latency, and computational overhead.
Identified Gaps and Challenges
Upon mapping over 20 existing systems onto the Experience Compression Spectrum, we discovered that each system operates at a fixed compression level. None of the analyzed systems support adaptive cross-level compression, which we refer to as the missing diagonal. This limitation indicates a significant area for advancement within the field.
Moreover, our study reveals that specialization in memory or skills alone does not suffice. Both communities are solving interrelated problems independently without sharing solutions, leading to duplicated efforts and missed opportunities for innovation.
Implications for Future Research
We advocate for a more integrated approach to knowledge lifecycle management, which remains largely unaddressed in the current literature. Key implications of our findings include:
- The need for cross-community collaboration to enhance knowledge transfer and innovation.
- A reconsideration of evaluation methods that account for varying compression levels across different systems.
- An exploration of the balance between transferability and specificity as it relates to compression.
Ultimately, addressing these challenges will pave the way for scalable, full-spectrum agent learning systems capable of leveraging the full range of experience compression.
Conclusion
The Experience Compression Spectrum offers a novel lens through which to view the development of LLM agents. By unifying memory, skills, and rules, we present a pathway for future advancements that can enhance the efficiency and effectiveness of AI systems. The ongoing dialogue between diverse communities will be crucial in overcoming existing barriers and fostering innovation in this rapidly evolving field.