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

Bing Zhu; Guanghui Wang; Peiyang He; Wei Qiu; Xing Zhang; Yanwei Cui; Ziyuan Li

REVIEW 4 major objections 3 minor 13 cited by

Memory, skills, and rules for LLM agents sit on one Experience Compression Spectrum, yet every system freezes at a fixed level.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 19:32 UTC pith:2L52BJPW

load-bearing objection Useful organizing frame for agent memory vs skills, but the compression bands and “missing diagonal” only count as findings if the ratio is operationalized—and we only have the abstract. the 4 major comments →

arxiv 2604.15877 v2 pith:2L52BJPW submitted 2026-04-17 cs.AI cs.CLcs.MA

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 This is my paper

classification cs.AI cs.CLcs.MA

keywords Experience Compression SpectrumLLM agentsagent memoryskill discoverydeclarative rulescompression ratiomissing diagonalknowledge lifecycle

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

As LLM agents scale to long-horizon multi-session work, managing accumulated experience becomes a bottleneck. Memory systems and skill discovery both extract reusable knowledge from traces, yet the communities barely cite each other. This paper introduces the Experience Compression Spectrum, placing episodic memory (roughly 5–20×), procedural skills (50–500×), and declarative rules (1,000×+) on a single axis of rising compression that directly cuts context size, retrieval latency, and compute. Mapping more than twenty systems shows every one operates at a fixed, predetermined level; none supports adaptive movement across levels—the missing diagonal. A sympathetic reader cares because the unification explains why current agents cannot scale efficiently and supplies design principles for full-spectrum systems that manage knowledge over its whole lifecycle.

Core claim

Memory, skills, and rules are points on a single Experience Compression Spectrum of increasing compression relative to raw interaction traces, and every surveyed agent system is locked to one fixed predetermined level on that spectrum rather than supporting adaptive cross-level compression.

What carries the argument

The Experience Compression Spectrum—an axis that ranks agent knowledge representations by compression ratio (5–20× episodic memory, 50–500× procedural skills, 1,000×+ declarative rules)—and uses that ranking both to unify previously separate literatures and to expose the absence of adaptive cross-level mechanisms.

Load-bearing premise

That heterogeneous memory, skill, and rule systems can be meaningfully ordered on one comparable numeric compression-ratio axis with the stated bands, so that “fixed level” and “missing diagonal” are empirical claims rather than metaphors.

What would settle it

Either the discovery of an existing agent that dynamically compresses the same experience into episodic, procedural, and declarative forms and selects among them at runtime, or direct measurement showing the claimed compression ratios do not hold for the mapped systems.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Agent architectures can be redesigned to move adaptively across compression levels instead of locking to one.
Shared sub-problems solved independently by the memory and skill communities become exchangeable once the spectrum is recognized.
Evaluation protocols must be decoupled from fixed compression levels if fair comparison across systems is desired.
Higher compression systematically trades specificity for transferability, guiding when each level should be preferred.
Knowledge lifecycle management becomes a first-class requirement for long-horizon agents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

An adaptive compression controller could emerge as a new architectural primitive, selecting the right spectrum level for each retrieval or planning step.
Without cross-level mechanisms, multi-session agents will continue to hit context and compute walls even as base models improve.
Compression ratio itself can serve as a quantitative bridge metric for papers that currently talk past one another.
Pipelines that automatically extract higher-compression rules from lower-compression skills or memories become a natural next research direction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Useful organizing frame for agent memory vs skills, but the compression bands and “missing diagonal” only count as findings if the ratio is operationalized—and we only have the abstract.

read the letter

The one thing to know: this is a position/survey paper that puts agent memory, skill libraries, and rules on one Experience Compression Spectrum, reports <1% cross-community citation over 1,136 references from 22 primary papers, and claims every mapped system sits at a fixed compression level—none does adaptive cross-level work (the “missing diagonal”).

What is actually new is the framing plus the bibliometric and gap claims. The ingredients—episodic stores, skill extraction, rule induction—are established. Credit where it is due: if the citation analysis holds, naming a near-total silo between memory and skills communities is useful, and the abstract’s further points (shared sub-problems reinvented independently, evaluation coupled to compression level, transferability vs specificity, neglected knowledge lifecycle) are the right questions for long-horizon agents. As a taxonomy that forces those communities into one picture, it can organize work that currently talks past itself.

The soft spot is real and matches the stress-test. The 5–20× / 50–500× / 1,000×+ bands and the “fixed level / missing diagonal” claim only work if a single comparable compression ratio is defined across traces, retrieved snippets, skill programs, and declarative rules. The abstract assigns the bands and reports the mapping without saying what the ratio is (tokens, cost, information rate, reuse judgment) or how mixed systems are scored. Without that protocol, the spectrum is a useful taxonomy and the missing diagonal is a metaphor, not an audited empirical gap. We also cannot check inclusion criteria for the 22 papers or the 1,136-reference set from the abstract alone. That is a methods gap, not a reason to dismiss the framing.

Who it is for: people building LLM agent memory, skill libraries, or multi-session systems who need a shared map of the design space. It deserves a serious referee—position papers that reorganize a siloed subfield should not be desk-rejected for being conceptual. I would not cite it yet until the compression definition and paper list are clear. Reading-group only if someone is actively designing multi-level agent learning; otherwise wait for the full PDF and skim the spectrum figure and citation claim.

Referee Report

4 major / 3 minor

Summary. The manuscript proposes the Experience Compression Spectrum as a unifying framework for LLM agent systems that manage accumulated experience. It positions episodic memory, procedural skills, and declarative rules as points on a single axis of increasing compression (stated bands: 5–20×, 50–500×, and 1,000×+ respectively), arguing that higher compression reduces context use, retrieval latency, and compute. From a citation analysis of 1,136 references across 22 primary papers (cross-community citation rate <1%) and a mapping of 20+ systems, the authors claim every surveyed system operates at a fixed, predetermined compression level and none supports adaptive cross-level compression—the “missing diagonal.” They further argue that specialization alone is insufficient, that evaluation methods couple tightly to compression level, that transferability rises with compression at the cost of specificity, and that knowledge lifecycle management is neglected, and they articulate open problems and design principles for full-spectrum agent learning.

Significance. If the spectrum is operationally well-defined and the mapping holds, the paper would usefully reframe two largely non-communicating communities (agent memory and skill discovery) under one axis, surface a concrete architectural gap (adaptive cross-level compression), and give designers a vocabulary for trading specificity against transfer and cost. The bibliometric observation of near-zero cross-citation and the explicit “missing diagonal” framing are potentially high-value contributions for a survey/position piece in cs.AI. Credit is due for attempting a parameter-light taxonomy with falsifiable-style empirical claims (fixed-level operation; no adaptive systems) rather than a purely narrative survey. Significance is conditional on the compression axis being measurable across heterogeneous artifacts.

major comments (4)

[Abstract (Experience Compression Spectrum; missing diagonal)] The central empirical claim—that 20+ systems each occupy a fixed compression level and that none supports adaptive cross-level compression—requires a single, comparable compression ratio defined across episodic traces, retrieved snippets, skill programs, and declarative rules. The abstract assigns numeric bands (5–20× / 50–500× / 1,000×+) and reports the mapping without stating the ratio’s definition (tokens-in/tokens-out, retrieval cost, information rate, human-judged reuse, or other), the measurement protocol, or how mixed-level systems are scored. Without that operationalization, “fixed level” and “missing diagonal” are not yet empirical results. This must be specified and applied consistently in the full mapping section.
[Abstract (citation analysis of 1,136 references / 22 primary papers)] The citation analysis (1,136 references, 22 primary papers, <1% cross-community rate) is load-bearing for the claim that the communities do not exchange solutions. Inclusion criteria for the 22 primaries, the partition into “memory” vs. “skills” communities, and the definition of a cross-community citation need to be stated so the rate is reproducible and selection bias can be assessed. If the primary set is small or the partition is post hoc, the <1% figure may overstate isolation.
[Abstract (mapping of 20+ systems; missing diagonal)] The claim that “every system operates at a fixed, predetermined compression level” and that “none supports adaptive cross-level compression” is absolute. The mapping must state how systems that combine memory stores with skill libraries or rule extractors were classified, and whether any system with even partial level-switching was considered and rejected. Absolute “none” claims are only as strong as the coverage and scoring rule of the 20+ system map.
[Abstract (evaluation / transferability / lifecycle claims)] Several secondary claims—evaluation methods tightly coupled to compression level; transferability increasing with compression at the cost of specificity; knowledge lifecycle management largely neglected—are asserted as shown results. Each needs an explicit evidence basis (which systems, which metrics, which lifecycle stages) so they do not rest only on the spectrum metaphor. If they are interpretive synthesis rather than measured findings, they should be labeled as such.

minor comments (3)

[Abstract] The compression bands are given as ranges (5–20×, 50–500×, 1,000×+) with no units or reference corpus; even a brief parenthetical on what “×” multiplies would help readers before the full methods section.
[Abstract] “Rules” / declarative knowledge should be briefly scoped (hand-written policies, extracted logical rules, natural-language principles, etc.) so the high-compression end of the spectrum is not ambiguous.
[Abstract] The phrase “specialization alone is insufficient (both communities independently solve shared sub-problems without exchanging solutions)” would benefit from one concrete shared sub-problem named in the abstract or early text for grounding.

Circularity Check

0 steps flagged

No significant circularity: abstract-only survey/position paper proposes a taxonomy and reports a bibliometric observation without defining quantities in terms of fitted parameters or self-citation chains that force the result.

full rationale

The available material is only the abstract of a survey/position paper. It proposes the Experience Compression Spectrum as a unifying framing that places memory, skills, and rules on one axis of increasing compression, reports a citation analysis of 1,136 references across 22 primary papers showing <1% cross-community citation, and maps 20+ systems to claim every system is fixed-level (the “missing diagonal”). None of these steps is a derivation that defines a quantity from a fitted parameter and then re-presents it as a prediction, nor does the abstract invoke a uniqueness theorem or ansatz from the authors’ prior work as an external fact. The numeric bands (5–20×, 50–500×, 1,000×+) and the mapping are presented as proposed taxonomy and observational claim, not as results forced by construction. Self-citation is not load-bearing in the abstract; residual selection bias in the 22 papers cannot be verified from the abstract alone and does not constitute circularity under the stated rules. Correctness risk (whether a single comparable compression ratio is operationalized) is outside the circularity pass. Score 0; steps empty.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

Abstract-only: free parameters are the stated compression bands and the choice of 22 primary papers / 20+ systems. Axioms are standard LLM-agent domain assumptions plus the ad-hoc claim that one compression axis is the right unifier. Invented entities are the spectrum construct and the “missing diagonal” gap label; neither has independent evidence beyond the authors’ mapping.

free parameters (4)

episodic_compression_band = 5–20×
Abstract asserts 5–20× for episodic memory without a measurement protocol visible here; the band is used to place systems on the spectrum.
procedural_compression_band = 50–500×
Abstract asserts 50–500× for procedural skills; used as a fixed placement on the unifying axis.
declarative_compression_band = 1000×+
Abstract asserts 1,000×+ for declarative rules; anchors the high-compression end of the spectrum.
primary_paper_set_size = 22 primary / 20+ systems
Choice of 22 primary papers and the 20+ systems mapped determines the citation-rate and missing-diagonal conclusions; selection criteria not given in abstract.

axioms (3)

domain assumption LLM agents accumulate interaction traces whose management is a primary bottleneck for long-horizon multi-session use.
Stated in the opening of the abstract; standard in agent-memory literature but not proved here.
ad hoc to paper Memory, skills, and rules can be ordered on a single comparable axis of experience compression.
Core modeling choice of the paper; enables the spectrum and missing-diagonal claims.
domain assumption Cross-community citation rate below 1% indicates insufficient exchange of solutions between memory and skill communities.
Bibliometric proxy for intellectual isolation; assumes citation rate tracks solution transfer.

invented entities (2)

Experience Compression Spectrum no independent evidence
purpose: Unify memory, skills, and rules as increasing compression of agent experience.
New organizing construct introduced by the paper; independent evidence would require a measurable compression metric validated outside this framing.
missing diagonal no independent evidence
purpose: Name the claimed gap that no system supports adaptive cross-level compression.
Gap label derived from the authors’ mapping; falsifiable only if the mapping and compression metric are public and reproducible.

pith-pipeline@v1.1.0-grok45 · 6147 in / 2905 out tokens · 28901 ms · 2026-07-12T19:32:34.911713+00:00 · methodology

0 comments

read the original 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 \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.

Figures

Figures reproduced from arXiv: 2604.15877 by Bing Zhu, Guanghui Wang, Peiyang He, Wei Qiu, Xing Zhang, Yanwei Cui, Ziyuan Li.

**Figure 1.** Figure 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. that implies agent systems should perform upward … view at source ↗

Review history (3 revisions) →

discussion (0)

Forward citations

Cited by 13 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents
cs.AI 2026-07 conditional novelty 7.0

An evaluation metric evolved from ten reference examples beat hidden unit tests on code generation and sufficed to drive a self-improving skill loop, while an unguarded version collapsed into an always-pass grader tha...
Co-Evolving Skill Generation and Policy Optimization
cs.CL 2026-06 unverdicted novelty 7.0

Framework estimates context-dependent marginal utility of candidate skills via reward gaps in matched base vs. skill-augmented rollouts to filter skills and co-train policy as generator.
TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory
cs.DB 2026-06 unverdicted novelty 7.0

TOKI types four common contradiction-resolution heuristics as bitemporal operators on a dual-row schema, supplies soundness theorems, and shows via a verdict matrix that it alone avoids three write-time anomalies whil...
Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries
cs.AI 2026-05 unverdicted novelty 7.0

Identifies library drift as a failure mode in self-evolving LLM skill libraries and shows a governance recipe improves pass@1 from 0.258 to 0.584 on MBPP+ hard-100.
Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries
cs.AI 2026-05 unverdicted novelty 7.0

The paper diagnoses library drift in self-evolving LLM skill libraries and demonstrates a governance recipe raising pass@1 from 0.258 to 0.584 on MBPP+ hard-100.
Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents
cs.AI 2026-07 conditional novelty 6.5

Double Ratchet co-evolves transparent metrics from small anchors with a skill lifecycle, recovering 88–110% of the lift that ground-truth or best rubrics would enable.
Ratchet: A Minimal Hygiene Recipe for Self-Evolving LLM Agents
cs.AI 2026-05 conditional novelty 6.0

Ratchet provides a minimal hygiene recipe for self-managing skill libraries in frozen LLM agents, delivering +0.328 rolling-mean pass@1 gain on MBPP+ hard-100 and +0.22 peak lift on SWE-bench Verified.
Ratchet: A Minimal Hygiene Recipe for Self-Evolving LLM Agents
cs.AI 2026-05 conditional novelty 6.0

Ratchet: outcome-driven retirement, a bounded active-cap, and a meta-skill authoring prior convert a +0.0pp self-authored skill library into a +0.33 rolling-mean gain on MBPP+ hard-100 with a frozen LLM.
Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries
cs.AI 2026-05 conditional novelty 6.0

A governance recipe—retire under-performing skills, cap the active set, and impose a meta-skill authoring style—raises held-out MBPP+ hard-100 pass@1 from 0.258 to 0.584, though the ungoverned 'drift' baseline itself ...
Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning
cs.LG 2026-05 unverdicted novelty 6.0

SLIM dynamically optimizes active external skills in agentic RL via leave-one-skill-out marginal contribution estimates and three lifecycle operations, outperforming baselines by 7.1% on ALFWorld and SearchQA while sh...
Trace2Policy: From Expert Behavior Traces to Self-Evolving Decision Agents
cs.AI 2026-06 unverdicted novelty 5.0

Trace2Policy's EISR iteratively refines expert-derived rules into compiled Python code reaching 79.6% accuracy on skewed compliance tasks, outperforming one-shot LLM distillation and a deployed LLM baseline.
SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution
cs.CL 2026-05 unverdicted novelty 5.0

SkillsVote is a governance system for agent skills that profiles corpora, recommends via search, and gates updates on successful reusable outcomes, yielding benchmark gains without model changes.
Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning
cs.LG 2026-05 unverdicted novelty 5.0

SLIM dynamically optimizes the active external skill set in agentic RL via leave-one-skill-out marginal contribution estimates and lifecycle operations, delivering a 7.1% average gain over baselines on ALFWorld and Se...