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

2026-07-31

AI Summary

This work addresses the "memory hysteresis" problem observed in self-evolving large language model agents under dynamic memory quota adjustments. To mitigate this issue, the authors propose CrystalMem, an elastic memory mechanism that introduces, for the first time, memory crystallization and verifiable recrystallization processes. CrystalMem employs a four-state fidelity model to dynamically degrade and restore memory entries, integrated with advantage-weighted influence ranking and dependency-coupling strategies to efficiently reconstruct capabilities under strict byte and computational constraints. Extensive experiments across seven environments, seventeen baselines, and six backbone models demonstrate that CrystalMem consistently outperforms existing approaches: under a 50% memory budget, it matches the performance of the strongest full-budget baseline and achieves an average gain of 4.6 percentage points under identical conditions, effectively eliminating the capability gap caused by memory hysteresis.

Abstract

Memory for self-evolving large language model (LLM) agents is often provisioned as if its byte budget only grows. Cloud platforms, however, adjust quotas with load and cost, and we show that capability does not follow the budget back up: after a squeeze-and-recover cycle, the agent settles below its pre-squeeze level, a gap we call memory hysteresis. The cause is structural. Deletion and one-way compression discard the material needed for later rebuilding, and we prove that any policy that only keeps or drops entries carries a residual-deficit floor. We propose CrystalMem (Crystallized Memory), an elastic memory sidecar that demotes entries across four fidelity states under a crystallization-energy schedule, orders demotions by advantage-weighted influence with dependency coupling, and recovers capability through verified recrystallization under explicit compute and byte caps. Across seven environments, seventeen methods, and six backbones, with multi-tenant serving and a physical edge-cloud deployment, CrystalMem achieves the highest restored capability in every setting and closes the loop left open by every baseline. From a 50% byte budget, CrystalMem matches the strongest budgeted baseline at full provision on every environment; at equal budgets, it leads by +4.6 pp on average.

Problem

• memory hysteresis
• elastic memory
• self-evolving LLM agents
• knowledge crystallization
• byte budget

Innovation

• elastic memory
• knowledge crystallization
• memory hysteresis
• LLM agents
• fidelity states