Source-linked AI summary
Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents
Xing Zhang, Guanghui Wang, Yanwei Cui, Wei Qiu, Ziyuan Li, Bing Zhu, Peiyang He
TL;DR
Persistent, long-horizon agents need better ways to manage accumulated interaction experience, while memory and skill-learning communities remain largely disconnected. The paper proposes the Experience Compression Spectrum, maps existing systems onto it, and identifies the missing diagonal: no mapped system adaptively selects compression levels. It concludes that the framework is promising but remains conceptual and requires experimental validation.
Problem
Long-horizon agents accumulate experience that strains context and retrieval budgets, while memory and skill communities address reusable experience with below-1% cross-community citation.
Method
The paper formalizes the Experience Compression Spectrum, maps 20+ systems across compression levels, and derives design principles and testable predictions for adaptive knowledge management.
Results
Existing systems cluster at Levels 1 and 2, with virtually no Level 3 systems and no adaptive cross-level compression, the missing diagonal.
Takeaways & Limitations
Compression increases generalizability and transferability while reducing context consumption and retrieval demands, but greater compression reduces context specificity.
Takeaways & Limitations
The framework is conceptual rather than empirically validated, and its four discrete levels simplify compression that may be continuous.
Abstract
from arXiv · showhide
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 \emph{Experience Compression Spectrum}, a unifying framework that positions memory, skills, and rules as points along a single axis of increasing compression (5--20$\times$ for episodic memory, 50--500$\times$ for procedural skills, 1{,}000$\times$+ for declarative rules), directly reducing context consumption, retrieval latency, and compute overhead. Mapping 20+ systems onto this spectrum reveals that every system operates at a fixed, predetermined compression level: none supports adaptive cross-level compression, a gap we term the \emph{missing diagonal}. We further show that specialization alone is insufficient (both communities independently solve shared sub-problems without exchanging solutions), that evaluation methods are tightly coupled to compression levels, that transferability increases with compression at the cost of specificity, and that knowledge lifecycle management remains largely neglected. We articulate open problems and design principles for scalable, full-spectrum agent learning systems.
1. Introduction
Persistent, long-horizon agents accumulate interaction experience faster than context windows and retrieval budgets can efficiently handle. The paper unifies memory, skills, and rules as increasingly compressed forms of reusable knowledge, while identifying severe separation between the communities developing them.
- Motivation: Long-horizon agents accumulate interaction traces that can overwhelm context windows and retrieval budgets, making experience management a scalability challenge.Memory systems extract experiential knowledge, while skill systems discover reusable procedural capabilities from execution traces.
- Community disconnect: A citation analysis of 1,136 references across 22 primary papers found cross-community citation rates below 1%.Memory papers cite skill work at 0.7% (4/566), while skill papers cite memory work at 1.2% (7/570).
- Experience Compression Spectrum: Memory extraction, skill discovery, and rule learning are framed as one operation: compressing interaction experience into reusable knowledge at different granularities.The paper characterizes memory, skills, and rules as approximately 10×, 100×, and 1,000×+ compression, respectively.
- Experience Compression Spectrum: The Experience Compression Spectrum positions memory, skills, and rules on a single axis from raw traces to abstract rules.The framework treats these forms as progressively more compressed representations of experience.
- Missing automation: Practitioners manually distill deployment experience into reusable rules, but existing systems operate at a single, predetermined compression level.The paper identifies this lack of automated movement across compression levels as a central gap.
2. The Experience Compression Spectrum
The Experience Compression Spectrum formalizes four levels that transform raw interaction traces into increasingly abstract, reusable knowledge. Mapping existing systems reveals clustering at episodic memory and procedural skill levels, little adaptive cross-level support, and substantial lifecycle and evaluation gaps.
- Formal Framework: The framework defines four compression levels, from raw traces through episodic memories and procedural skills to declarative rules.The levels progress from 1:1 raw records to increasingly abstract and reusable artifacts.
- Properties of the Spectrum: L1 episodic memory compresses traces roughly 5–20×, L2 procedural skills 50–500×, and L3 declarative rules 1,000×+.These levels correspond to progressively greater reuse and lower context or retrieval overhead.
- Properties of the Spectrum: Higher compression increases generalizability while reducing context-specificity and information retention.The spectrum also trades acquisition cost against maintenance cost, making abstraction increasingly reusable but less specific.
- Mapping Existing Systems: Mapping 20+ systems shows ten systems clustered at L1 and eight at L2, while L3 automation remains absent.L1 systems converge on structured episodic records, whereas L2 systems distill traces into reusable behavioral patterns.
- Mapping Existing Systems: Existing performance evidence favors higher-compression representations, but results span different benchmarks and cannot be directly compared across rows.The reported comparisons include SkillRL’s +68.5pp over L1 trajectory retrieval and curated skills’ +16.2pp benefit, while self-generated skills provide +0.0pp.
- The Missing Diagonal: The missing diagonal is the absence of systems that adapt compression level, promote knowledge upward, or demote it when abstraction becomes too broad.Current systems use predetermined output levels; adaptive selection is framed as learning what kind of knowledge to extract.
3. What the Spectrum Reveals
The spectrum exposes shared challenges across memory and skill communities, including evaluation mismatch, transferability–specificity trade-offs, and neglected lifecycle management. It also motivates unified infrastructure and testable predictions for adaptive compression.
- Specialization Alone Is Insufficient: Memory and skill communities independently address shared retrieval, conflict, staleness, and downstream-utility evaluation problems without exchanging solutions.
- Specialization Alone Is Insufficient: A unified architecture could share retrieval, conflict-resolution, and lifecycle infrastructure across level-specific compressors.
- Transferability Increases with Compression Level: L1 memories transfer across base models, while L2 skills transfer across tasks and model sizes, indicating increasing transferability with compression.
- Transferability Increases with Compression Level: +68.5pp higher success is reported for L2 skills than L1 trajectory retrieval on ALFWorld.
- Lifecycle Management Remains an Afterthought: Knowledge maintenance remains incomplete: systems lack full-lifecycle support for conflict detection, staleness recognition, deprecation, and cross-level consistency.
- Testable Predictions: The framework predicts that L2 compression should outperform L1 retrieval for cross-domain transfer and that multi-level stores should outperform either level alone.
4. Open Problems and Design Principles
The paper identifies adaptive level selection, cross-level consistency, and lifecycle governance as pressing problems, and proposes level-agnostic, bidirectional, and metadata-supported designs. It also describes idle-time consolidation and a concrete promotion/demotion architecture, while noting that the framework remains conceptual.
- Problem 1: Adaptive level selection: Adaptive level selection requires a meta-controller that chooses compression level or levels for each new trace.
- Problem 2: Cross-level consistency: Cross-level consistency requires detecting and resolving inconsistencies when multiple knowledge levels describe the same behavior.
- Additional open directions: Open directions include reward-free compression, cross-domain skill-transfer protocols, and extension to multimodal traces.
- Design principles: The proposed design principles are a level-agnostic compression core and bidirectional promotion and demotion between memory, skill, and rule levels.
- Problem 3: Principled lifecycle management: Lifecycle governance should attach provenance, confidence, usage frequency, and validation time to each knowledge artifact.
- Concrete architecture: Idle-time consolidation is proposed for upward compression, but no current system implements it and existing systems compress synchronously at ingestion.
- Concrete architecture: A meta-controller can route traces by novelty and frequency, promote repeated patterns upward, and demote invalid generalizations after validation failure.
- Limitations: The framework is conceptual rather than empirically validated, uses four discrete levels as a simplifying abstraction, and surveys a Q1 2026 snapshot.
5. Conclusion
The paper unifies memory, skill discovery, and rule learning as points on an Experience Compression Spectrum. Mapping existing systems reveals a missing diagonal and several structural challenges for scalable agent learning.
- The Experience Compression Spectrum places agent memory, skill discovery, and rule learning on a single compression axis.
- Mapping 20+ systems exposes missing adaptive cross-level compression, level-coupled evaluation, a transferability–specificity trade-off, and neglected lifecycle management.
LLM/Agent Usage Disclosure
The authors state that LLMs polished prose and improved clarity, while the intellectual contributions remain theirs.
- LLMs were used for prose polishing and clarity improvement, not for the paper’s intellectual contributions.