Framework unifies memory, skills, and rules in LLM agent systems
Researchers Xing Zhang, Guanghui Wang, and Yanwei Cui propose the Experience Compression Spectrum, a unifying framework that positions memory, skills, and rules along a single axis of increasing compression to address fragmentation between LLM agent memory and skill-discovery research communities.
Score breakdown
Practitioners building long-running LLM agents can use this framework to identify which compression level their memory or skill system targets and design toward adaptive, cross-level compression to reduce context costs and avoid redundant engineering work already solved in adjacent communities.
- 01Proposed by Xing Zhang, Guanghui Wang, and Yanwei Cui as the Experience Compression Spectrum framework.
- 02Citation analysis of 1,136 references across 22 primary papers found cross-community citation rate below 1%.
- 03Episodic memory achieves 5–20× compression; procedural skills 50–500×; declarative rules 1,000×+.
Xing Zhang, Guanghui Wang, and Yanwei Cui present the Experience Compression Spectrum, a unifying conceptual framework for LLM agent learning systems. The core insight is that episodic memory, procedural skills, and declarative rules — typically studied in separate research communities — can all be understood as forms of experience compression operating at different ratios: 5–20× for episodic memory, 50–500× for procedural skills, and 1,000×+ for declarative rules. By framing these as points on a single axis, the framework directly addresses context consumption, retrieval latency, and compute overhead as agents scale to long-horizon, multi-session deployments.
The authors conclude by articulating open problems and design principles for building scalable, full-spectrum agent learning systems.
To motivate the unification, the authors conducted a citation analysis of 1,136 references across 22 primary papers, finding a cross-community citation rate below 1% between the agent memory and agent skill-discovery communities. Mapping 20+ existing systems onto the spectrum exposes a structural gap the authors call the "missing diagonal": every surveyed system operates at a fixed, predetermined compression level, and none supports adaptive compression across levels. The paper further identifies that both communities independently solve shared sub-problems without exchanging solutions, that evaluation methods are tightly coupled to specific compression levels, that transferability increases with compression at the cost of specificity, and that knowledge lifecycle management is largely neglected. The authors conclude by articulating open problems and design principles for building scalable, full-spectrum agent learning systems.
Key facts
- 01Proposed by Xing Zhang, Guanghui Wang, and Yanwei Cui as the Experience Compression Spectrum framework.
- 02Citation analysis of 1,136 references across 22 primary papers found cross-community citation rate below 1%.
- 03Episodic memory achieves 5–20× compression; procedural skills 50–500×; declarative rules 1,000×+.
- 04Mapping 20+ systems onto the spectrum reveals all operate at fixed, predetermined compression levels.
- 05No existing system supports adaptive cross-level compression — a gap the authors call the 'missing diagonal'.
- 06Transferability increases with compression level, but at the cost of specificity.
- 07Knowledge lifecycle management is identified as a largely neglected area across both communities.