Memory-Augmented Agents

Memory-augmented agents study how AI systems can retain and use information beyond the current prompt or observation, helping them handle long tasks, partial observability, and changing user or world states. Approaches range from episodic, semantic, and procedural stores—often organized as graphs, event records, or compressed visual and action histories—to learned policies that decide what to write, retrieve, update, or forget. Work spans embodied control and navigation, GUI and web use, dialogue, coding, and research, with evaluations including RoboMemArena, RoboMME, GOAT-Bench, LongMemEval, and software-engineering benchmarks. A central tension is that more memory can improve transfer and continuity but also introduce stale or misleading evidence, interference, cost, privacy leaks, and persistent attacks, making selective retrieval and longitudinal safety evaluation important open problems.