Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution

TL;DR

The paper tackles long-horizon GUI automation under context window limits by introducing Darwinian Memory System (DMS), a training-free, self-evolving memory that decomposes workflows into reusable action subsequences and evolves memory via survival pressure, mutation, and Bayesian risk assessment. DMS combines a Planner-Actor framework with dual-factor memory retrieval, in-place mutations, and adaptive pruning to prevent context pollution and adapt to changing GUIs. Across the AndroidWorld benchmark, DMS substantially increases success rate and execution stability while reducing latency, supporting scalable deployment on general-purpose MLLMs without fine-tuning. The work demonstrates a biologically inspired lifelong-learning paradigm for GUI agents by transforming historical experience into an evolving, high-signal memory ecosystem that improves robustness and efficiency over time.

Abstract

Multimodal Large Language Model (MLLM) agents facilitate Graphical User Interface (GUI) automation but struggle with long-horizon, cross-application tasks due to limited context windows. While memory systems provide a viable solution, existing paradigms struggle to adapt to dynamic GUI environments, suffering from a granularity mismatch between high-level intent and low-level execution, and context pollution where the static accumulation of outdated experiences drives agents into hallucination. To address these bottlenecks, we propose the Darwinian Memory System (DMS), a self-evolving architecture that constructs memory as a dynamic ecosystem governed by the law of survival of the fittest. DMS decomposes complex trajectories into independent, reusable units for compositional flexibility, and implements Utility-driven Natural Selection to track survival value, actively pruning suboptimal paths and inhibiting high-risk plans. This evolutionary pressure compels the agent to derive superior strategies. Extensive experiments on real-world multi-app benchmarks validate that DMS boosts general-purpose MLLMs without training costs or architectural overhead, achieving average gains of 18.0% in success rate and 33.9% in execution stability, while reducing task latency, establishing it as an effective self-evolving memory system for GUI tasks.

Figures (11)

• Figure 1: Performance Overview. In multi-app GUI scenarios, DMS consistently boosts Accuracy and Stability across diverse general-purpose models while reducing task latency.
• Figure 2: Overview of the Darwinian Memory System. The architecture integrates a Planner-Actor framework with a self-evolving memory mechanism. The Feedback Regulation module (Top) suppresses low-reputation plans via Bayesian risk assessment to drive exploration. The Memory Bank (Center) supports In-place Evolutionary Updates, where mutated trajectories with higher efficiency overwrite existing entries. To ensure system stability, Homeostatic Regulation (Bottom) dynamically prunes obsolete memories using the Elbow Method based on a multi-factor survival value (i.e., utility, decay, and reliability).
• Figure 3: Memory construction overview. Unlike short-step methods requiring expensive post-hoc summarization (a), our approach (b) leverages long-horizon plans as natural, ground-truth summaries for atomic actions.
• Figure 4: Heatmap of Memory Reuse Rates. The visualization displays the memory reuse probability across 5 execution rounds for three models. Darker colors indicate higher reuse rates.
• Figure 5: Performance and Stability Analysis. (a) Success rate comparison of various open-source models with and without DMS across multiple experimental trials. The lower panel provides a granular breakdown of performance across varying task difficulties. (b) Stability landscape visualizing the Success Rate Retention (SRR) of different methods, distinguished by color. Please refer to Appendix \ref{['app:SRR']} for the formal definition of SRR.