SE-GA
SE-GA: Memory-Augmented Self-Evolution for GUI Agents
Shilong Jin, Lanjun Wang, Zhuosheng Zhang · 2026-05-16
- Management
- T2 Dynamic Memory
- Memory changes during deployment through a stable management policy.
- Representation
- Hybrid
- Distinct operative representation paths are deliberately linked or jointly consumed by downstream reasoning, generation, or control.
How experience becomes memory
Each path is one memory component, traced from what is stored to how it is found again.
episodic, semantic, and experiential GUI memory
- Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
- Stored inHierarchical multi-tier store, Vector-indexed collection
- Retrieved bySimilarity search, Planner / tool-assisted retrieval
- Source evidenceSource not retained
TTME retrieves hierarchical episodic, semantic, and experiential memory at inference.Full paper: 2.2. Memory Mechanisms · PDF p. 2
LoRA-adapted GUI policy conditioned on accumulated interaction experience
- Represented asExperience-conditioned parametric memoryInternalized latent & parametric
- Stored inModular adapter bank
- Retrieved byLatent / parametric activation
- Source evidenceSource not retained
MASE uses accumulated GUI interaction experience to adapt the agent policy with LoRA, internalizing reusable experience in parameters.Full paper: Model · PDF p. 8
How memory changes (T2 operations)
Update · Rule-based update
During operation, SE-GA applies fixed-rule insertion or replacement to episodic, semantic, and experiential GUI memory; the memory state changes while the governing update policy remains fixed.
Forgetting · Discrete eviction
During operation, SE-GA applies explicit deletion or eviction of stored items to episodic, semantic, and experiential GUI memory; the memory state changes while the governing update policy remains fixed.