Long‑term memory is emerging as a fundamental substrate for personalized AI, yet most existing systems treat personalization as a set of static records embedded in a global latent space and accessed via a single similarity metric. This creates a mismatch with the reality of data mining, where evidence arrives as a temporal event stream while the dominant abstraction remains a searchable record collection. We propose to model long‑horizon personalization as a user‑specific dynamical state space endowed with locally heterogeneous geometry. In this context, geometry is a computational language that captures stable versus volatile regions, variable‑rate drift, heterogeneous neighborhoods, and uncertainty about the current user state. Profiles and isolated events stay useful as points, but interaction, feedback, and elapsed time generate trajectories within the space. Memory access thus becomes trajectory‑conditioned reconstruction of the relevant user state rather than a simple nearest‑neighbor lookup.
Review