CrystalMem: Elastic Memory for Self-Evolving LLM Agents
A recent paper published on arXiv (2608.00303) presents CrystalMem, an innovative elastic memory sidecar aimed at mitigating memory hysteresis in self-evolving large language model (LLM) agents. The study reveals that when cloud services modify memory allocations based on demand and expenses, agents that experience a squeeze-and-recover process often perform below their original capacity, a situation referred to as memory hysteresis. This issue arises from structural factors: the loss of data during deletion and one-way compression prevents effective rebuilding. The authors demonstrate that any policy focused solely on retaining or discarding entries results in a persistent deficit. CrystalMem utilizes a crystallization-energy schedule to manage demotions across four fidelity states and enhances recovery through verified recrystallization within specified compute and byte limits. The research was conducted across seven environments, utilizing seventeen methods and six backbone models. The paper was authored by a team of researchers and made available on arXiv on August 1, 2026.
Key facts
- CrystalMem is an elastic memory sidecar for LLM agents.
- It addresses memory hysteresis, where agents settle below pre-squeeze capability after quota recovery.
- The paper proves a residual-deficit floor for policies that only keep or drop entries.
- CrystalMem uses four fidelity states and a crystallization-energy schedule.
- Demotions are ordered by advantage-weighted influence with dependency coupling.
- Recovery is achieved through verified recrystallization under compute and byte caps.
- Evaluated across seven environments, seventeen methods, and six backbone models.
- Paper available on arXiv with ID 2608.00303.
Entities
Institutions
- arXiv