ElasticMem: Latent Memory as a Learnable Resource for LLM Agents

In three linesElasticMem introduces a learnable latent memory framework for LLM agents with adaptive retrieval and elastic budget allocation via learned policy. On Qwen2.5-3B and 7B backbones, achieves 26.2% and 24.6% QA accuracy gains, 66.3% and 27.2% ALFWorld success improvements, with lowest token cost.

Summary generated by Claude — human-verified