Beyond Training: Enabling Self-Evolution of Agents with MOBIMEM

TL;DR

MOBIMEM addresses the challenge of post-deployment self-evolution for LLM-based agents by decoupling evolution from model weights through a memory-centric design. It introduces three specialized memories—DisGraph-based Profile Memory, multi-level Experience Memory, and Action Memory with ActTree/ActChain—plus OS-inspired services for scheduling, record-and-replay, and interruption handling. The approach yields substantial gains in profile alignment, task success, and end-to-end latency on mobile devices, and enables scalable, cross-app execution through fine-grained parallelism. By enabling continuous improvement via agent memories and safe replay, MOBIMEM demonstrates a practical path toward persistent, edge-friendly GUI agents without costly retraining. The results suggest that memory-centric architectures can significantly advance personalization, capability growth, and efficiency in real-world mobile AI workflows.

Abstract

Large Language Model (LLM) agents are increasingly deployed to automate complex workflows in mobile and desktop environments. However, current model-centric agent architectures struggle to self-evolve post-deployment: improving personalization, capability, and efficiency typically requires continuous model retraining/fine-tuning, which incurs prohibitive computational overheads and suffers from an inherent trade-off between model accuracy and inference efficiency. To enable iterative self-evolution without model retraining, we propose MOBIMEM, a memory-centric agent system. MOBIMEM first introduces three specialized memory primitives to decouple agent evolution from model weights: (1) Profile Memory uses a lightweight distance-graph (DisGraph) structure to align with user preferences, resolving the accuracy-latency trade-off in user profile retrieval; (2) Experience Memory employs multi-level templates to instantiate execution logic for new tasks, ensuring capability generalization; and (3) Action Memory records fine-grained interaction sequences, reducing the reliance on expensive model inference. Building upon this memory architecture, MOBIMEM further integrates a suite of OS-inspired services to orchestrate execution: a scheduler that coordinates parallel sub-task execution and memory operations; an agent record-and-replay (AgentRR) mechanism that enables safe and efficient action reuse; and a context-aware exception handling that ensures graceful recovery from user interruptions and runtime errors. Evaluation on AndroidWorld and top-50 apps shows that MOBIMEM achieves 83.1% profile alignment with 23.83 ms retrieval time (280x faster than GraphRAG baselines), improves task success rates by up to 50.3%, and reduces end-to-end latency by up to 9x on mobile devices.

Figures (13)

• Figure 1: MobiMem tames the trade-off between AI agents' latency and accuracy by a memory-centric design.
• Figure 2: Three-layer architecture of MobiMem: specialized multi-agent layer (top), agent-tailored memory layer (middle), and OS integration layer (bottom).
• Figure 3: Architecture of the Profile Memory module, showing profile updating, storage, and retrieval workflows.
• Figure 4: Experience template example showing template structure and experience generation/retrieval workflows.
• Figure 5: Two structures of Action Memory: ActTree (left) for prefix reuse and ActChain (right) for prefix-suffix reuse.