Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents
April 17, 2026
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\(\times\) 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 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↩︎
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 [1]–[3], while the agent skill community builds frameworks for discovering and reusing procedural capabilities from execution traces [4]–[6].
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 [7], [8] cites any memory system, and only one memory survey [9] 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 [10], 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 (\({\sim}10\times\) compression); a skill system extracts reusable behavioral patterns (\({\sim}100\times\)); a rule system extracts abstract decision principles (\({\sim}1{,}000\times\)+). 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 [11] 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 [12] maps onto our \(L_1\)/\(L_2\) boundary, and Fitts & Posner’s [13] 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 (\(L_0 \to L_3\)). 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 (2); (2) map 20+ systems onto this spectrum, revealing that none supports adaptive cross-level compression, the “missing diagonal” (2); (3) quantify the community disconnect (\(<\)1% cross-citation) and expose four structural insights invisible from within either community (3); and (4) articulate open problems and design principles for scalable, full-spectrum agent learning (4).
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 1 (Interaction Trace). An interaction trace \(\mathcal{T} = \{(s_t, a_t, o_t, f_t)\}_{t=1}^{N}\) is a sequence of states \(s_t\), actions \(a_t\), observations \(o_t\), and feedback signals \(f_t\) collected during agent execution.
Definition 2 (Experience Compression Function). An experience compression function \(C_L: \mathcal{T} \rightarrow \mathcal{K}_L\) maps traces to knowledge artifacts at compression level \(L \in \{0,1,2,3\}\).
The four levels are:
2.1.0.1 Level 0: Raw Trace.
The uncompressed interaction record. Format: complete logs, execution trajectories. Compression ratio: 1:1. Reusability: minimal, entirely context-bound.
2.1.0.2 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: \({\sim}5\)–\(20\times\). Reusability: low to moderate, tied to specific episodes.
2.1.0.3 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: \({\sim}50\)–\(500\times\). Reusability: high, transferable across similar situations.
2.1.0.4 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: \({\sim}1000\times{+}\). 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 (\(L_0 \to L_3\)), 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 [14] compresses multi-session conversation history (\({\sim}26{,}000\) tokens) into retrieved memory entries (\({\sim}1{,}800\) tokens), roughly \(15\times\) at \(L_1\). Trace2Skill [15] distills traces from 200 tasks via 128 parallel sub-agents into a compact skill directory, roughly \(100\)–\(500\times\) at \(L_2\).
Acquisition cost vs.maintenance cost. \(L_1\) memories are cheap to acquire (single trace) but expensive to maintain at scale: an \(L_1\)-only system accumulating thousands of entries per day will exhaust any practical retrieval budget within weeks. \(L_3\) 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 \(L_1\)-only agent storing 1,000 episodes at \({\sim}500\) tokens each maintains a \({\sim}500\)K-token knowledge store that must be indexed and searched at every decision. Compressing to \(L_2\) skills reduces this to \({\sim}5\)K tokens; to \(L_3\) rules, \({\sim}500\) tokens: a \(100\)–\(1{,}000\times\) 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 \(L_0 \to L_3\), 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 (1, 2), revealing that each operates at a fixed, predetermined compression level.1
2.3.0.1 Level 1 (Episodic Memory).
Ten systems cluster here (2), spanning diverse mechanisms (LLM-driven extraction [14], agentic indexing [17], hierarchical storage [2], RL-optimized memory operations [18]–[21], meta-learned memory architectures [22], multi-agent coordination [23], and governance with temporal decay [24]) but converging on the same output: structured episodic records.
2.3.0.2 Level 2 (Procedural Skill).
Eight systems cluster here. Voyager [4] pioneered the paradigm; CASCADE [25] 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 (1). SkillRL [26] reports \(+68.5\)pp over \(L_1\) trajectory retrieval on ALFWorld; Trace2Skill [15] outperforms human-written skills by \(+21.5\)pp on SpreadsheetBench. SkillsBench [27] 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.
2.3.0.3 Cross-level (\(L_1 \leftrightarrow L_2\)).
ExpeL [28] and AutoAgent [29] operate at two levels simultaneously, but both use predetermined levels without adaptive selection: two-speed systems, not continuous spectra.
2.3.0.4 Level 3 (Declarative Rule).
Notably sparse. No surveyed system automates rule extraction from agent experience. Constitutional AI [30] uses pre-specified rules; reward design methods (PBRS [31], process rewards [32], rule-based rewards [33]) encode \(L_3\)-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 \(L_3\) extraction are technical: (i) distinguishing causal regularities from incidental correlations is harder than episode extraction; (ii) rules without \(L_2\) 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 false sense that evaluation is solved. However, recent empirical work provides initial \(L_3\) evidence: RuleShaping [34] studies 25,000+ natural-language rules for coding agents and finds that negative constraints (guardrails) improve performance by 7–14pp while positive directives hurt, suggesting that \(L_3\) compression must preserve shaping signals rather than prescriptive instructions. SEVerA [35] adds formal verification to self-evolving agents, orthogonal to compression level but critical for correctness. Closing this loop at \(L_3\), verbal reinforcement learning with explicit insight governance [36] pairs rule extraction with a persistent evidence log and a curation policy, showing that the same accumulated experience either degrades performance below a zero-shot baseline or improves it depending on whether the governance loop is present.
| Study | Comparison | Bench. | \(\Delta\) |
|---|---|---|---|
| SkillRL | \(L_2\) vs.\(L_1\) retrieval | ALFWorld | \(+68.5\)pp |
| Trace2Skill | \(L_2\) vs.human skill | SpreadSh. | \(+21.5\)pp |
| Trace2Skill | \(L_2\) vs.no skill | SpreadSh. | \(+42.1\)pp |
| SkillsBench | curated \(L_2\) vs.none | multi-task | \(+16.2\)pp |
| SkillsBench | self-gen \(L_2\) vs.none | multi-task | \(+0.0\)pp |
| EvoSkill | \(L_2\) vs.no skill | BrowseC. | \(+5.3\)% |
| RuleShaping | \(L_3\) constr.vs.-shot | SWE-b. | \(+7\)–\(14\)pp |
| System | Level | Mechanism | Lifecycle |
|---|---|---|---|
| MemSkill | \(L_1\) | RL | partial |
| Mem0 | \(L_1\) | LLM | partial |
| A-MEM / MemoryOS | \(L_1\) | LLM | |
| Memory-R1 / Mem-\(\alpha\) | \(L_1\) | RL | partial |
| MemPO | \(L_1\) | RL | |
| ALMA | \(L_1\) | meta-learn | |
| MemMA | \(L_1\) | multi-agent | partial |
| SSGM | \(L_1\) | governance | partial |
| Voyager | \(L_2\) | LLM | |
| SkillWeaver | \(L_2\) | LLM | |
| EvoSkill | \(L_2\) | search | partial |
| CASCADE | \(L_2\) | LLM | |
| AutoSkill | \(L_2\) | LLM | ✔ |
| Trace2Skill | \(L_2\) | LLM | |
| SkillRL | \(L_2\) | RL | |
| EvolveR | \(L_2\) | RL | partial |
| ExpeL | \(L_1\)–\(L_2\) | LLM | partial |
| AutoAgent | \(L_1\)–\(L_2\) | LLM | partial |
| SEVerA | orthogonal | verify |
2.4 The Missing Diagonal↩︎
The mapping reveals a striking pattern: systems cluster at Level 1 or Level 2, with minimal cross-level work and virtually no Level 3 (1). We call this the missing diagonal: the absence of systems that can:
Adaptively select the appropriate compression level \(L\) for a given trace \(\mathcal{T}\).
Promote knowledge upward (\(C_{L \to L'}\) for \(L' > L\)) when sufficient evidence accumulates (many memories \(\to\) one skill \(\to\) one rule).
Demote knowledge downward (\(C_{L \to L'}\) for \(L' < L\)) when a rule proves too abstract for a specific context.
This gap reflects a problem framing limitation (each community defines its output format a priori), not missing engineering effort. Adaptive level selection is a meta-learning problem: learning what kind of knowledge to extract, not just what knowledge to extract. It is also a scalability bottleneck: an \(L_1\)-only system accumulating episodic memories faces linear growth in storage and retrieval cost, eventually degrading performance as irrelevant entries dilute useful knowledge. Upward compression (\(L_1 \to L_2 \to L_3\)) is the natural solution: consolidate recurring patterns into compact skills or rules that scale sub-linearly with experience volume.
3 What the Spectrum Reveals↩︎
Viewing existing work through the compression lens exposes four structural insights invisible from within either community alone.
3.1 Specialization Alone Is Insufficient↩︎
One might argue that the memory–skill separation reflects healthy specialization: memory systems serve personalization, skill systems serve capability. But the spectrum reveals that both communities independently solve the same sub-problems (retrieval over growing knowledge stores, conflict detection between contradictory entries, staleness recognition, and evaluation of downstream utility) without sharing solutions. Moreover, real deployments require both personalization and capability: an agent that abstracts reusable workflows but forgets user preferences is incomplete; one that remembers context but cannot consolidate recurring patterns will overwhelm its retrieval budget (2.4). The spectrum makes these shared challenges visible, suggesting that a unified architecture, with level-specific compressors sharing common retrieval, conflict-resolution, and lifecycle infrastructure, would avoid redundant engineering across communities.
3.2 Evaluation Methods Are Level-Coupled↩︎
\(L_1\) systems evaluate via QA metrics (F1, exact match), \(L_2\) via task success rate, and \(L_3\) has no established methodology. This coupling means systems are optimized for their level’s metric, which may not reflect true downstream utility: MemSkill achieves strong QA F1 but its impact on multi-step task completion is unclear; EvoSkill’s task success metric cannot capture user-specific context. A unified evaluation framework should assess knowledge artifacts by their downstream impact on agent performance across time, regardless of compression level: an efficiency metric that captures value per token of stored knowledge.
3.3 Transferability Increases with Compression Level↩︎
Empirical evidence supports a monotonic relationship: \(L_1\) memories transfer across base models (MemSkill: LLaMA \(\to\) Qwen); \(L_2\) skills transfer across tasks (EvoSkill: SealQA \(\to\) BrowseComp, \(+5.3\%\)) and model sizes (Trace2Skill: 35B \(\to\) 122B, \(+57.7\) pp) [15]. This relationship likely follows a concave curve, with the sweet spot at \(L_2\) balancing transferability with specificity. A controlled experiment holding source experience constant while varying only compression level has not been conducted, but existing evidence is consistent (1): SkillRL reports \(+68.5\)pp higher success with \(L_2\) skills than \(L_1\) trajectory retrieval on ALFWorld [26].
3.4 Lifecycle Management Remains an Afterthought↩︎
Existing systems focus on knowledge acquisition while treating maintenance as secondary. Some include partial mechanisms (Mem0’s LLM-driven operations, Memory-R1’s RL-learned updates, ExpeL’s importance counting, AutoSkill’s [37] version management) but these remain isolated. No system addresses the full lifecycle: conflict detection across knowledge types, staleness recognition, principled deprecation, and cross-level consistency. Merging skills does not always help (including success-trace signals can cause \(\pm\)21pp performance swings [15]), and agent behavior can degrade over extended operation even when correctness is maintained. A recent diagnosis of skill-library drift [38] shows that without outcome-driven retirement and a bounded active-cap, self-evolving libraries degrade below their no-skill baseline, isolating lifecycle management as the failure’s root cause rather than skill authoring. The software engineering community’s version control and deprecation practices offer a concrete opportunity for transfer to knowledge artifacts.
3.4.0.1 Testable predictions.
The spectrum generates predictions that distinguish it from pure taxonomy: (i) \(L_2\) compression should outperform \(L_1\) retrieval on cross-domain transfer when holding source experience constant; (ii) a multi-level knowledge store (\(L_1\)+\(L_2\)) should outperform either level alone, with the gap widening as deployment length increases; (iii) the transferability–specificity curve should be concave, with \(L_2\) at the sweet spot; and (iv) \(L_3\) rules should help most when framed as constraints rather than directives, consistent with early evidence from RuleShaping [34]. Validating or falsifying these predictions would establish whether the spectrum is a useful design tool or merely a descriptive framework.
4 Open Problems and Design Principles↩︎
The unified perspective exposes concrete research problems, three of which we consider most pressing for building scalable agent learning systems.
4.0.0.1 Problem 1: Adaptive level selection.
Design a meta-controller that, given a new trace \(\mathcal{T}\), determines the optimal compression level(s) \(L^*\). Recent agentic RL surveys [39] find that RL primarily amplifies existing capabilities; genuinely new strategies emerge only under stringent conditions, suggesting that the level of compression may matter more than the algorithm. This requires a value-of-information framework: how much future utility does \(\mathcal{K}_{L}\) provide relative to the cost of computing and maintaining it? A key design tension is whether to jointly optimize the meta-controller and level-specific compressors, or to fix the meta-controller and optimize only the compressors. Analysis of iterative generative optimization [40] shows that design decisions (starting artifacts, credit horizons, experience batching) often dominate algorithm choice, suggesting that fixing the meta-strategy while optimizing components may be more practical.
4.0.0.2 Problem 2: Cross-level consistency.
When knowledge exists at multiple levels simultaneously (e.g., an \(L_1\) memory and an \(L_3\) rule about the same behavior), how should the system detect and resolve inconsistencies? The analogy to database normalization is suggestive but incomplete: knowledge artifacts are semantic, not relational.
4.0.0.3 Problem 3: Principled lifecycle management.
Borrow from software engineering: version control for knowledge artifacts, dependency tracking between levels, deprecation protocols, and conflict resolution. Recent workflow optimization surveys [41] advocate a “minimum plasticity” principle; an analogous principle for knowledge management, promoting only when evidence warrants, would reduce maintenance burden while preserving adaptability.
Additional open directions include reward-free compression (extending self-training methods like STaR [42] and ReST [43] to guide multi-level compression without clean reward signals), cross-domain transfer protocols (can \(L_2\) skills from one deployment warm-start another?), and extension to multimodal traces (visual observations and cross-modal interactions amplify the need for efficient compression).
We propose three design principles for systems that address these problems:
Level-agnostic compression core. The compression engine should be parameterized by output level, not hard-coded. DSPy [44] demonstrates that declarative specifications can be compiled into optimized pipelines at a predetermined level; a level-agnostic core would generalize this.
Bidirectional promotion/demotion. Knowledge should flow upward (memory \(\to\) skill \(\to\) rule when patterns accumulate) and downward (rule \(\to\) skill \(\to\) memory when context demands specificity).
Continuous lifecycle governance. Every knowledge artifact should carry metadata (provenance, confidence, usage frequency, last validation time) enabling principled maintenance.
The spectrum also suggests idle-time consolidation: upward compression (\(L_1 \to L_2 \to L_3\)) should occur during low-activity periods, analogous to hippocampal-neocortical consolidation during sleep [11]. No current system implements this; all compress synchronously at ingestion time.
Concretely, a “diagonal” system might comprise three components. A meta-controller routes each trace to one or more level-specific compressors based on novelty and frequency: a first-seen pattern stays at \(L_1\); once \(k\) similar \(L_1\) entries accumulate, the promotion engine consolidates them into an \(L_2\) skill; after cross-domain recurrence, the engine generalizes to an \(L_3\) rule. A promotion/demotion engine handles these transitions: newly promoted \(L_2\) skills are validated against held-out tasks before replacing their source \(L_1\) entries; demotion triggers when a skill is repeatedly retrieved but unused, or when an abstract rule fails in a specific context, reverting to fresh \(L_1\) evidence collection. A lifecycle manager tracks provenance, usage frequency, and last validation time for every artifact, enabling principled deprecation.
As a concrete example: a customer-support agent encounters a timeout on /api/export and stores an \(L_1\) memory. After five similar episodes, promotion produces an \(L_2\)
skill (Handle_Export_Timeout: check batch size, reduce if \({>}1000\) rows, retry). After dozens of instances across endpoints, generalization yields an \(L_3\) rule (“Timeouts on data-intensive endpoints typically stem from oversized batches”). If the rule fails in a novel context, demotion drops back to \(L_1\), restarting evidence collection.
Existing components (MemSkill’s RL controller for routing, EvoSkill’s Pareto selection for promotion, AutoSkill’s versioning [37] for
lifecycle) could serve as building blocks for this architecture.
4.0.0.4 Limitations.
Our framework focuses on scaffold-level knowledge; training-time methods (RLHF, Constitutional AI) are complementary but distinct. The framework is conceptual rather than empirically validated: the spectrum’s utility as a design tool awaits experimental confirmation. The four discrete levels are a simplifying abstraction; in practice, compression may be continuous (EvolveR’s [45] strategies straddle \(L_2\)/\(L_3\)). Our survey is a snapshot of Q1 2026; the field moves rapidly. While we focus on text-based agents, the spectrum naturally extends to multimodal experience: visual observations and cross-modal interactions amplify the need for efficient compression at every level.
5 Conclusion↩︎
We have proposed the Experience Compression Spectrum, revealing agent memory, skill discovery, and rule learning as points on a single compression axis. Mapping 20+ systems onto this spectrum exposed the missing diagonal (the absence of adaptive cross-level compression) alongside the insufficiency of specialization alone, level-coupled evaluation, the transferability–specificity trade-off, and neglected lifecycle management. With 10+ papers in Q1 2026 alone and production systems increasingly managing persistent experiential knowledge [46], this unification is timely: without a shared framework, two communities risk solving the same problem twice. The most promising direction is building systems that adaptively select compression granularity to match the value and generality of each experience, enabling agents that scale efficiently across long deployments.
Impact Statement↩︎
This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.
LLM/Agent Usage Disclosure↩︎
LLMs were used to polish prose and improve clarity of writing. All intellectual contributions (framework design, system analysis, citation analysis, and conclusions) are the authors’ own.