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

Researchers propose the Experience Compression Spectrum, a unifying framework that reconciles two separate research communities studying LLM agent memory and skill discovery by positioning them along a single compression axis. The framework identifies a critical gap—no existing system supports adaptive cross-level compression—and reveals that memory systems and skill discovery communities operate in isolation despite solving overlapping problems.

The paper addresses a fundamental scalability challenge in deploying LLM agents at production scale. As agents handle longer tasks across multiple sessions, their accumulated experience consumes exponentially more context tokens, increasing both computational overhead and latency. The research reveals a striking fragmentation: memory systems researchers and skill discovery researchers cite each other less than 1% of the time despite tackling nearly identical problems from different angles.

The Experience Compression Spectrum unifies these approaches by quantifying compression ratios across three knowledge types—episodic memory (5–20×), procedural skills (50–500×), and declarative rules (1,000×+). This hierarchy directly correlates with reduced context consumption and computational cost. However, the framework exposes a critical vulnerability: every existing system operates at a fixed, predetermined compression level with no mechanism to dynamically adjust between levels based on task requirements. This rigidity creates inefficiency.

For practitioners building AI agents, this framework provides actionable design principles for architecting more efficient systems. The finding that specialization alone is insufficient suggests that future competitive advantage lies in systems that can fluidly compress and decompress knowledge across the spectrum. The research also highlights that evaluation methodologies have become tightly coupled to specific compression levels, making cross-system comparisons difficult and potentially masking hybrid approaches.

The most pressing implication concerns knowledge lifecycle management, which the paper identifies as largely neglected. As agents accumulate experience over months or years, without proper lifecycle strategies, irrelevant or outdated knowledge will bloat context windows. Future agent platforms will likely need sophisticated curation systems.

• →Memory and skill discovery research communities operate in near-complete isolation despite solving shared sub-problems, indicating significant fragmentation in agent AI development.
• →No current LLM agent system supports adaptive cross-level knowledge compression, representing a major architectural gap for production deployments.
• →Compression ratios span three orders of magnitude (episodic to declarative), with higher compression improving transferability but reducing task-specific accuracy.
• →Evaluation methods are tightly coupled to compression levels, making it difficult to compare systems across the spectrum or discover hybrid approaches.
• →Knowledge lifecycle management remains largely unsolved, creating a critical bottleneck for long-horizon, multi-session agent deployments at scale.