LifeMem: Enabling Lifelong Experience Reuse for LLM Agents
Abstract Overview
LifeMem is a lifelong learning framework for LLM agents designed to accumulate and reuse past experience across heterogeneous environments while mitigating catastrophic forgetting. The framework groups accumulated interaction trajectories into workflow-based clusters and distills each cluster into reusable, high-level textual skills, retrieving both environment-aligned exemplars and abstract skills during inference. The authors evaluate LifeMem across 10 environments and over 13k tasks spanning embodied action, tool use, web search, data analysis, and web browsing, contributing more than 2k newly annotated interaction trajectories. Experimental results demonstrate that structured memory evolution enhances experience reuse, reduces forgetting on previously learned tasks, and improves transfer to unseen domains.
Novelty
LifeMem introduces a memory mechanism that organizes trajectories into functional workflow clusters rather than relying purely on surface semantic similarity, distilling skills locally within each cluster to avoid cross-environment noise. It also incorporates an adaptive cluster refinement mechanism governed by a geometric directional diversity criterion (Mean Resultant Length).
Results
Across three backbone LLMs, LifeMem achieves the highest average overall performance in the main lifelong learning benchmark: 44.13 for GPT-4o-mini, 48.02 for Deepseek-v3.2-exp, and 46.52 for Qwen3-32b, consistently outperforming ReAct, Synapse, ExpeL, and AutoSkill. It is also the only tested framework to maintain a positive average backward transfer score across all three backbones (+5.66, +3.11, and +6.68, respectively). Furthermore, LifeMem improves zero-shot transfer on unseen tasks such as FEVER and airline booking, while a buffer-based streaming strategy recovers substantial performance lost under randomized task sequences.
Key Points
- LifeMem organizes agent experience into workflow clusters and distills them into environment-invariant skills, combining them with environment-specific exemplars at inference.
- The evaluation spans 10 heterogeneous environments and over 13,000 tasks, supported by more than 2,000 newly annotated interaction trajectories for datasets lacking training splits.
- The method consistently achieves positive backward transfer across backbones, prevents negative interference from cluttered memory, and outperforms baselines in zero-shot cross-task transfer.