Research map / Multi-agent systems and games / LLM agents
Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Darwinian Memory System (DMS), a training-free memory for GUI agents that stores decomposed sub-task trajectories and prunes/penalizes them via survival value and Bayesian risk, boosting Planner-Actor MLLMs on AndroidWorld.
Problem
MLLM GUI agents struggle on long-horizon, cross-application tasks; existing memory systems use monolithic, static storage that becomes rigid under GUI changes and accumulates outdated or toxic experiences.
Approach
Decompose trajectories into reusable precondition-goal sub-task memory units, retrieve them with a multiplicative dual-factor similarity, and regulate the memory ecosystem with a survival value (marginal utility, sigmoid temporal decay, reliability penalty), elbow-based pruning, epsilon-mutation with in-place replacement, and Beta-Binomial risk gating.
Key results
- DMS-PA-Lite with Qwen2.5-VL-72B reaches 66.4% success rate on AndroidWorld vs 41.0% for the PA-Lite baseline (+25.4), with the best baseline UI-Venus at 65.9%.
- Stability (SRR) for Qwen3-VL-30B rises from 36.5% to 84.7% and GLM-4.5V from 53.0% to 87.0% with DMS.
- Removing the Standard goal-only key drops performance to 29.7%, below the memory-free baseline of 41.0%.
Surprising findings
- A goal-only retrieval key yields 29.7% success, below the memory-free PA-Lite baseline (41.0%), which the authors attribute to negative transfer from mismatched trajectories.
- Removing self-regulation yields a 258% larger disk footprint but the authors note its memory growth rate is lower than DMS's, attributed to obsolete memories blocking updates.
Limitations
- Baseline PA-Lite is reported at peak (best-of-N) performance in the main table while DMS is reported at final round, a comparison asymmetry; average-based results show smaller gains for some models.
- Efficacy is bounded for tasks needing high-granularity atomic actions, where single-step memories are filtered out and reuse stays at zero.
- Hyperparameters may need per-model tuning; evaluation limited to AndroidWorld.
Experimental setup
Datasets: AndroidWorld
Benchmarks: AndroidWorld
Metrics: success rate success retention rate memory reuse rate latency
Baselines: GPT-4o UI-TARS UI-Venus EchoTrail-GUI MobileUse AndroidGen Seed1.5-VL PA-Lite
Builds on: ReAct MemEvolve EchoTrail-GUI
Closest papers by meaning
- SE-GA: Memory-Augmented Self-Evolution for GUI Agents (ICML 2026 · similarity 0.92)
- MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.91)
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.91)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.89)
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.