Scaling Self-Evolving Agents via Parametric Memory

In three linesTMEM introduces a self-evolving parametric memory framework for LLM agents. Instead of storing experience solely as textual context, the agent absorbs distilled supervision into lightweight LoRA weights (Δ_t) via online updates, genuinely altering behavior within a single episode. Evaluated on LoCoMo, LongMemEval-S, and CL-Bench, TMEM consistently outperforms summary-based and retrieval-based baselines.

Summary generated by Claude — human-verified