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

A team of researchers led by Xing Zhang has published a seminal paper introducing the 'Experience Compression Spectrum,' a novel framework designed to unify the fragmented fields of memory systems and skill discovery for LLM agents. The core insight is that all methods for reusing agent experience—from detailed episodic memory to abstract declarative rules—exist on a single continuum defined by their compression ratio. The paper maps over 20 existing systems onto this spectrum, revealing compression ranges from 5–20x for episodic memory, 50–500x for procedural skills, and over 1,000x for declarative rules. This direct compression reduces critical bottlenecks like context consumption, retrieval latency, and compute overhead.

A striking finding from their citation analysis of 1,136 references across 22 key papers is a cross-community citation rate below 1%, indicating that memory and skill research silos are solving shared problems in isolation. The framework exposes a major architectural gap termed the 'missing diagonal': every current system operates at a fixed compression level, with none capable of adaptive, cross-level compression. The authors argue this limits agents' ability to balance transferability (which increases with compression) against specificity (which decreases). They conclude by outlining open problems and design principles necessary for building scalable, full-spectrum agent learning systems that can efficiently manage knowledge over long-horizon deployments.