Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language
This work is significant for adaptive agents in dynamic environments, particularly in fields like healthcare and customer service, where the ability to control what a model remembers is crucial.
The paper addresses the lack of user control in neural memory models, which typically assume a fixed objective and homogeneous information. They propose a generalized neural memory system that allows flexible updates based on natural language learning instructions, enabling selective learning from diverse information sources.
Modern machine learning models are deployed in diverse, non-stationary environments where they must continually adapt to new tasks and evolving knowledge. Continual fine-tuning and in-context learning are costly and brittle, whereas neural memory methods promise lightweight updates with minimal forgetting. However, existing neural memory models typically assume a single fixed objective and homogeneous information streams, leaving users with no control over what the model remembers or ignores over time. To address this challenge, we propose a generalized neural memory system that performs flexible updates based on learning instructions specified in natural language. Our approach enables adaptive agents to learn selectively from heterogeneous information sources, supporting settings, such as healthcare and customer service, where fixed-objective memory updates are insufficient.