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

# Unifying Memory, Skills, and Rules in LLM Agents Source: https://arxiv.org/html/2604.15877 ## Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents ###### Abstract 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 *Experience Compression Spectrum*, a unifying framework that positions memory, skills, and rules as points along a single axis of increasing compression (5–20× for episodic memory, 50–500× for procedural skills, 1,000×+ 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 *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. LLM Agents, Experience Compression, Agent Memory, Skill Discovery, Efficient Agentic Systems, Scalable Learning Level 0Raw Trace1:1Level 1Episodic Memory5–20×Level 2Procedural Skill50–500×Level 3Declarative Rule1000×+extractabstractgeneralizeConversation logsExecution tracesMemSkill, Mem0A-MEM, MemoryOSALMA, MemMAMemory-R1, Mem-αMemPO, SSGMVoyager, SkillWeaverEvoSkill, Trace2SkillAutoSkill, SkillRLCASCADE, EvolveR*(largely empty)*Constitutional AI(pre-specified only)missing diagonalExpeL, AutoAgent(cross-level, but fixed)Generalizability⟶⟵SpecificityFigure 1: The Experience Compression Spectrum. Existing agent learning systems map onto a single axis from raw traces to abstract rules. Memory systems cluster at Level 1, skill systems at Level 2, with Level 3 largely empty. A small number of cross-level systems (dashed) bridge Levels 1–2 but none support adaptive level selection. Compression ratios are approximate.## 1 Introduction As LLM agents move from single-session demos to persistent, long-horizon deployments, they accumulate vast interaction experience. An agent handling thousands of tasks per day generates traces that quickly overwhelm any practical context window or retrieval budget, making efficient experience management a first-order scalability challenge. Two research communities have emerged to address this: the agent memory community develops systems for extracting and retrieving experiential knowledge (Hu et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib10); Kang et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib13); Packer et al., 2023 (https://arxiv.org/html/2604.15877#bib.bib25)), while the agent skill community builds frameworks for discovering and reusing procedural capabilities from execution traces (Wang et al., 2023 (https://arxiv.org/html/2604.15877#bib.bib27); Zheng et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib44); Alzubi et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib1)). Despite addressing the same fundamental problem—extracting reusable knowledge from interaction experience—these communities are remarkably disconnected. A citation analysis of 1,136 references across 22 primary papers reveals a cross-community citation rate below 1%. Memory papers cite skill work at 0.7% (4/566); skill papers cite memory work at 1.2% (7/570). Neither skill survey (Jiang et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib12); Xu & Yan, 2026 (https://arxiv.org/html/2604.15877#bib.bib33)) cites any memory system, and only one memory survey (Yang et al., 2026a (https://arxiv.org/html/2604.15877#bib.bib36)) cites any skill work (only Voyager). This separation reflects a conceptual gap that limits the design of scalable agent systems. Scope. We study knowledge extracted at the *scaffold level*—runtime systems outside model weights. Training-time methods (RLHF (Ouyang et al., 2022 (https://arxiv.org/html/2604.15877#bib.bib24)), Constitutional AI) are complementary but out of scope. Our key observation is that memory extraction and skill discovery are instantiations of the same operation: *compressing interaction experience into reusable knowledge* at different granularities. A memory system extracts structured event records (∼10× compression); a skill system extracts reusable behavioral patterns (∼100×); a rule system extracts abstract decision principles (∼1,000×+). These are not three separate problems—they are three points on a single experience compression spectrum, where higher compression directly translates to reduced context consumption, faster retrieval, and lower compute overhead per decision. This observation has cognitive science analogs. Complementary Learning Systems (CLS) theory (McClelland et al., 1995 (https://arxiv.org/html/2604.15877#bib.bib20)) describes how the hippocampus rapidly encodes episodic memories that are gradually consolidated into neocortical knowledge—a biological compression spectrum that implies agent systems should perform upward compression during idle time. ACT-R’s declarative–procedural distinction (Anderson, 1983 (https://arxiv.org/html/2604.15877#bib.bib2)) maps onto our L1/L2 boundary, and Fitts & Posner’s (1967 (https://arxiv.org/html/2604.15877#bib.bib8)) skill acquisition theory shows that knowledge flows *bidirectionally*—explicit rules compile into automatic procedures through practice—a property no current system supports. Practitioners already perform full-spectrum compression manually: hundreds of thousands of developers maintain CLAUDE.md and .cursorrules files that distill deployment experience into reusable rules (L0→L3). Yet no system automates this process; each operates at a single, predetermined compression level. Contributions. We (1) formalize the Experience Compression Spectrum, unifying agent memory, skills, and rules on a single compression axis (Section 2 (https://arxiv.org/html/2604.15877#S2)); (2) map 20+ systems onto this spectrum, revealing that none supports adaptive cross-level compression—the “missing diagonal” (Section 2 (https://arxiv.org/html/2604.15877#S2)); (3) quantify the community disconnect (<1% cross-citation) and expose four structural insights invisible from within either community (Section 3 (https://arxiv.org/html/2604.15877#S3)); and (4) articulate open problems and design principles for scalable, full-spectrum agent learning (Section 4 (https://arxiv.org/html/2604.15877#S4)). ## 2 The Experience Compression Spectrum We formalize a four-level spectrum characterizing how interaction experience is progressively compressed into increasingly abstract and reusable knowledge. ### 2.1 Formal Framework ###### Definition 2.1 (Interaction Trace). An interaction trace T = {(s_t, a_t, o_t, f_t)}_{t=1}^N is a sequence of states s_t, actions a_t, observations o_t, and feedback signals f_t collected during agent execution. ###### Definition 2.2 (Experience Compression Function). An experience compression function C_L: T → K_L maps traces to knowledge artifacts at compression level L ∈ {0, 1, 2, 3}. The four levels are: #### Level 0 — Raw Trace. The uncompressed interaction record. Format: complete logs, execution trajectories. Compression ratio: 1:1. Reusability: minimal—entirely context-bound. #### Level 1 — Episodic Memory. Structured extraction of *what happened*, preserving key contextual details while discarding redundant interaction mechanics. Format: key-value pairs, timestamped event summaries (e.g., “[2026-03-15] User requested Q3 revenue analysis via SQL. Preferred tabular format.”). Compression ratio: ∼5–20×. Reusability: low to moderate—tied to specific episodes. #### Level 2 — Procedural Skill. Extraction of *how to act* in a class of situations, abstracting across instances into reusable behavioral patterns. Format: structured routines, code snippets, workflow templates (e.g., “Data_Analysis: (1) Confirm source, (2) Select tool, (3) Present in preferred format, (4) Verify.”). Compression ratio: ∼50–500×. Reusability: high—transferable across similar situations. #### Level 3 — Declarative Rule. Extraction of *what principles govern decisions*—domain-invariant knowledge. Format: natural language principles, constraints, policies (e.g., “Always verify computed results against source data before presenting.”). Compression ratio: ∼1000×+. Reusability: highest—domain-general, but may lack actionable specificity. ### 2.2 Properties of the Spectrum The four levels exhibit systematic trade-offs along three dimensions, each with direct efficiency implications. Generalizability vs. specificity. As compression increases (L0→L3), knowledge becomes more broadly applicable but less context-specific. Compression ratio vs. information retention. Higher levels discard more contextual detail via *semantic abstraction*: identifying which patterns generalize and which are incidental. As concrete examples: Mem0 (Chhikara et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib6)) compresses multi-session conversation history (∼26,000 tokens) into retrieved memory entries (∼1,800 tokens)—roughly 15× at L1. Trace2Skill (Ni et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib22)) distills traces from 200 tasks via 128 parallel sub-agents into a compact skill directory—roughly 100–500× at L2. Acquisition cost vs. maintenance cost. L1 memories are cheap to acquire (single trace) but expensive to maintain at scale—an L1-only system accumulating thousands of entries per day will exhaust any practical retrieval budget within weeks. L3 rules require many traces to induce but form a compact, low-maintenance set. This trade-off makes upward compression not merely desirable but *necessary* for agents that operate at scale over extended deployments. To quantify: an L1-only agent storing 1,000 episodes at ∼500 tokens each maintains a ∼500K-token knowledge store that must be indexed and searched at every decision. Compressing to L2 skills reduces this to ∼5K tokens; to L3 rules, ∼500 tokens—a 100–1,000× reduction in storage and retrieval overhead that compounds across thousands of daily decisions. Critically, these are not sequential pipeline stages. A system could compress directly from L0→L3, or maintain knowledge at multiple levels simultaneously. The spectrum describes the *output space*, not a fixed processing order. ### 2.3 Mapping Existing Systems We select systems that (a) learn from interaction traces (excluding pre-specified rules), (b) produce persistent knowledge artifacts, and (c) have been published since 2023. We position 20+ systems on the spectrum (Figure 1 (https://arxiv.org/html/2604.15877#S0.F1), Table 2 (https://arxiv.org/html/2604.15877#S2.T2)), revealing that each operates at a fixed, predetermined compression level.¹¹Our selection is not exhaustive; additional L1 systems such as LightMem (Fang et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib7)) further corroborate the clustering pattern. #### Level 1 (Episodic Memory). Ten systems cluster here (Table 2 (https://arxiv.org/html/2604.15877#S2.T2)), spanning diverse mechanisms—LLM-driven extraction (Chhikara et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib6)), agentic indexing (Xu et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib34)), hierarchical storage (Kang et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib13)), RL-optimized memory operations (Zhang et al., 2026b (https://arxiv.org/html/2604.15877#bib.bib41); Yan et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib35); Wang et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib29); Li et al., 2026a (https://arxiv.org/html/2604.15877#bib.bib16)), meta-learned memory architectures (Xiong et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib32)), multi-agent coordination (Lin et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib19)), and governance with temporal decay (Lam et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib15))—but converging on the same output: structured episodic records. #### Level 2 (Procedural Skill). Eight systems cluster here. Voyager (Wang et al., 2023 (https://arxiv.org/html/2604.15877#bib.bib27)) pioneered the paradigm; CASCADE (Huang et al., 2025 (https://arxiv.org/html/2604.15877#bib.bib11)) chains cumulative skill creation with autonomous evolution. An emerging consensus reinforces our framework: *distilling traces into abstract skills consistently outperforms storing them in retrieval memory banks* (Table 1 (https://arxiv.org/html/2604.15877#S2.T1)). SkillRL (Xia et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib31)) reports +68.5 pp over L1 trajectory retrieval on ALFWorld; Trace2Skill (Ni et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib22)) outperforms human-written skills by +21.5 pp on SpreadsheetBench. SkillsBench (Li et al., 2026b (https://arxiv.org/html/2604.15877#bib.bib17)) further finds that curated skills help (+16.2 pp) while LLM-self-generated skills provide no benefit (+0.0 pp)—compression level alone is insufficient; the *fidelity* of the compression process determines whether the artifact is useful or merely compact noise. #### Cross-level (L1↔L2). ExpeL (Zhao et al., 2024 (https://arxiv.org/html/2604.15877#bib.bib43)) and AutoAgent (Wang et al., 2026 (https://arxiv.org/html/2604.15877#bib.bib28)) operate at two levels simultaneously, but both use *predetermined* levels without adaptive selection—two-speed systems, not continuous spectra. #### Level 3 (Declarative Rule). Notably sparse. No surveyed system automates rule extraction from agent experience. Constitutional AI (Bai et al., 2022 (https://arxiv.org/html/2604.15877#bib.bib3)) uses *pre-specified* rules; reward design methods (PBRS (Ng et al., 1999 (https://arxiv.org/html/2604.15877#bib.bib21)), process rewards (Lightman et al., 2023 (https://arxiv.org/html/2604.15877#bib.bib18)), rule-based rewards (Shao et al., 2024 (https://arxiv.org/html/2604.15877#bib.bib26))) encode L3-type knowledge but are human-designed, not learned. Weight-level rules (via RLHF) are static after training, opaque to inspection, and cannot be updated without retraining. Scaffold-level rules would be inspectable, editable, and deployable without gradient updates. The barriers to automated L3 extraction are technical: (i) distinguishing causal regularities from incidental correlations is harder than episode extraction; (ii) rules without L2 grounding risk being too abstract; (iii) no methodology exists for evaluating rule quality (infinite regress of meta-evaluation); and (iv) LLM-as-Judge gives a fals