Experience Compression Spectrum: an architectural framework unifying memory, skills, and rules in LLM agents

What is the Experience Compression Spectrum?

A new arXiv preprint published on 17 April 2026 proposes a unified theoretical framework for how LLM agents handle experience across long-horizon, multi-season deployments. Rather than treating memory, skills, and rules as separate architectural components, authors Xing Zhang and colleagues position them along a single axis of increasing compression:

• Episodic memory — raw records of what happened, compression 5–20×
• Procedural skills — parameterised routines learned from patterns, 50–500×
• Declarative rules — general statements that hold across contexts, 1000×+

The idea is simple but powerful: all three are different levels of the same process — compressing experience into reusable knowledge. The only difference is how much context is lost for the sake of greater conciseness.

What does the analysis of existing systems reveal?

The authors identify three systemic problems:

1. Fixed compression level. Most agents operate at one point on the spectrum — some remember everything, others extract rules. But real experience is not uniform — some things deserve detailed memory (edge cases), others deserve extreme compression (stable procedures). Systems without adaptive flexibility pay a price at one end or the other.

2. Memory and skills do not communicate. Research communities working on memory (long-term context, RAG, episode replay) and skills (skill learning, program synthesis) do not exchange results. The authors argue these are fundamentally the same thing — compression of experience — just developed in silos.

3. Evaluation differs by level. How do you measure “good memory” vs. “good skill” vs. “good rule”? Each level has its own benchmarks, making it difficult to compare systems operating at different points on the spectrum.

What are the design principles for full-spectrum agents?

The paper does not propose a concrete implementation but rather principles for agents operating across the full spectrum:

• Dynamic positioning — the agent itself chooses the compression level for each experience, based on pattern frequency and confidence
• Bidirectional movement — skills can be distilled from memory, rules from skills; but also in reverse: when a rule breaks down, the agent must be able to “decompress” back to episodic detail
• Lifecycle management — rules and skills become stale as context changes; experience needs a revision mechanism, not just accumulation

Why does this matter?

Long-horizon agents — those operating for weeks or months in the same context (customer support, tech support, personal assistants, coding) — need experience. But pure memorisation does not scale (the context window grows, costs grow), and premature compression loses information. The paper argues that compression is a spectrum, not a binary choice, and that the next generation of agents should be designed with that in mind.

Implications for builders

For teams building production agents, the message is architectural: instead of separate modules for memory, skills, and rules, consider a single memory hierarchy with mechanisms for promotion and demotion across levels. Mechanisms such as summarisation pipelines, skill extractors, and rule inducers are parts of the same system — they just operate at different compression levels.

The paper is a preprint without experimental results from new models — it is more of a position paper that defines a common vocabulary for the field. But that is precisely its value: teams currently building long-horizon agents can use it as a guide when designing their memory architecture.