Inception: Aligning Agent and Human Memory Through Consolidation for Always-on AI Wearables
In revision, 2026.
Overview
Always-on AI wearables need to connect individual observations with the contexts and routines that make them useful. Inception investigates how memory consolidation can support this connection. Drawing on Complementary Learning Systems theory, it organizes memory into three functional layers: episodic memory for particular events, semantic memory for recurring contexts, and schema memory for routines.
The architecture progressively compiles observations into reusable situational and behavioral knowledge. It supports two forms of assistance: detecting exceptions to familiar routines and transferring learned knowledge to a new context. For example, recognizing a departure routine can make a phone left on a counter relevant to the current activity.
We examined these capabilities through an automated ablation and a controlled study with 12 participants using recorded routine scenarios. The full architecture improved performance on routine-exception cases, and participants preferred its exception-detection responses over a flat-memory baseline. Transfer preferences were evenly split. These results suggest where structured consolidation can help proactive assistance, while leaving live, long-term use and the management of changing routines for further study.