Memory Papers
Memory Architecture

CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization

AuthorsBeining Wu, Jun Huang

arXiv 20262026
arXivPDF

TL;DR

CrystalMem uses a four-state reversible fidelity ladder with influence-based scheduling and verified recrystallization to close memory hysteresis loops, recovering to 57.4 avg capability vs 52.8 for R3Mem (+4.6 pp).

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Elastic memory budgets cause memory hysteresis with up to 17.6 pp residual deficit

Under squeeze and recover cycles, capability does not return with the bytes, leaving a residual deficit up to 17.6 pp for LRU.

Self-evolving LLM agents with deletion or one-way compression settle below pre-squeeze capability, so memory hysteresis and residual deficit directly limit long-term agent capability.

HOW IT WORKS

CrystalMem — Crystallized memory with reversible fidelity states

CrystalMem wraps the memory plane with Monitor, Crystallize, and Recrystallize, moving entries across a four-state fidelity ladder using advantage-weighted influence and dependency coupling.

Conceptually, CrystalMem treats memory like a multi-level storage hierarchy where hot items stay in fast full form and colder ones crystallize into compact residues, ready to be reheated later.

This crystallization plus verified recrystallization lets CrystalMem trade bytes for compute and recover capability that a plain context window or binary retention would permanently lose.

DIAGRAM

Elastic-stage lifecycle of CrystalMem under a squeeze–recover budget

This diagram shows how CrystalMem runs Monitor, Crystallize, and Recrystallize across the seven-stage elastic budget cycle.

DIAGRAM

Evaluation pipeline across seven environments and seventeen methods

This diagram shows how the experiments evaluate CrystalMem against baselines over the elastic budget schedule.

PROCESS

How CrystalMem Handles an Elastic-Stage Lifecycle

  1. 01

    Phase I: Monitor

    CrystalMem uses Monitor to update advantage-weighted influence Ie and coupled values ˜ve from retrieval logs and graded outcomes at each stage.

  2. 02

    Phase II: Crystallize

    CrystalMem runs Crystallize to solve the crystallization-energy schedule, demoting entries along the four-state fidelity ladder until the byte budget is satisfied.

  3. 03

    Phase III: Recrystallize

    CrystalMem selects a promotion set under compute and byte caps, regenerates entries from residues, and applies the verification gate before lifting fidelity.

  4. 04

    Elastic Budget Cycle

    CrystalMem repeats Monitor, Crystallize, and Recrystallize over the seven-stage budget path, closing the capability loop and minimizing hysteresis area.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Memory hysteresis identification and analysis

    CrystalMem formalizes memory hysteresis with a loop-area metric, showing LRU can leave a 17.6 pp residual deficit and proving a residual-deficit floor for binary retention policies.

  • 02

    CrystalMem elastic memory sidecar

    CrystalMem introduces a four-state fidelity ladder with crystallization energy, advantage-weighted influence, dependency coupling, and verified recrystallization under explicit compute and byte caps.

  • 03

    Comprehensive experimental evaluation

    CrystalMem is evaluated on seven environments, seventeen methods, and six backbones, achieving 57.4 average restored capability vs 52.8 for R3Mem and reducing mean hysteresis area to 0.013.

RESULTS

By the Numbers

Avg restored capability

57.4

+4.6 pp over R3Mem

Mean hysteresis area

0.013

−0.031 vs R3Mem

Mean residual deficit

0.5 pp

order of magnitude below binary baselines

Iso-capability budget

50%

matches CURATOR at full budget on six of seven environments

Across SF, TTL, LoCoMo, LME, ALFWorld, WebArena, and SynDrift, CrystalMem is compared under the same elastic budget schedule. The 57.4 vs 52.8 average restored capability shows CrystalMem closes the hysteresis loop better than R3Mem while using the same byte budget.

BENCHMARK

By the Numbers

Across SF, TTL, LoCoMo, LME, ALFWorld, WebArena, and SynDrift, CrystalMem is compared under the same elastic budget schedule. The 57.4 vs 52.8 average restored capability shows CrystalMem closes the hysteresis loop better than R3Mem while using the same byte budget.

BENCHMARK

Iso-budget comparison over one elastic cycle (seven-environment average)

Average restored capability across seven environments at the end of the elastic cycle.

KEY INSIGHT

The Counterintuitive Finding

After the budget returns to 100%, strong baselines like CURATOR still lose 6.6 pp, while CrystalMem leaves only a 0.5 pp residual deficit.

This is surprising because elastic cloud platforms usually assume capability tracks provision, but memory hysteresis shows that naive deletion or one-way compression permanently damages agent performance.

WHY IT MATTERS

What this unlocks for the field

CrystalMem makes it possible to run self-evolving LLM agents under elastic memory budgets without permanently sacrificing capability after squeezes.

Builders can now treat memory as a reversible, compute-tradable resource, confidently sharing and resizing stores across tenants while preserving long-term agent competence.

~14 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization

Answers use this explainer on Memory Papers.

Checking…