Memory-Augmented Language Model Agents

This field studies how language-model agents can retain and reuse information beyond the current prompt, including user facts, task state, prior successes and failures, and procedural skills. Common approaches include retrieval from vector or graph stores, hierarchical summaries and context compression, and learned policies for deciding what to write, retrieve, update, or forget; some methods also encode experience into model parameters or latent representations. Evaluations span conversational recall benchmarks such as LoCoMo and LongMemEval, interactive settings such as ALFWorld and WebShop, and applications in coding, robotics, and clinical reasoning. A central tension is that memory can improve continuity and reduce context costs but also introduce stale or misleading information, negative transfer, privacy risks, and security vulnerabilities, so reliable lifecycle management and realistic long-horizon evaluation remain open challenges.