A new research paper wants to replace AI assistants' static memory banks with something closer to a living map.
Most personalized AI systems store long-term memory as discrete records sitting in one uniform latent space, all searched with the same similarity measure. The paper argues that setup is a poor fit for how people actually interact with a system: real user history is a stream of events over time, not a static database. Instead, it proposes modeling memory as a user-specific state space with geometry that varies by region - some areas stable, others prone to fast drift, each with its own uncertainty about where the user currently stands. Under this approach, pulling up a memory means reconstructing the relevant state from a trajectory, not just finding the nearest neighbor.
That reframing matters because it shifts the job of personalization from cataloging what a user has said to tracking how a user changes, including how confident a system should be that an old data point still holds. It is also a direct critique of the record-and-embed approach that most memory-augmented AI products already ship with.
For now this is a proposed framework, not a working product, so the real test is whether anyone builds a memory system this way and it holds up once it leaves the paper.