UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams
UniMem routes new tasks to episodic storage and recurring patterns into expandable model memory.
The paper targets LLM agents working through task streams where task boundaries and labels are not given. UniMem uses learnable routing tokens to decide whether to rely on retrieval from an episodic buffer or consolidate reliable patterns into parametric memory blocks. The authors say this avoids fixed parameter budgets and uncontrolled growth while reducing dependence on inference-time retrieval. In long-horizon streaming experiments, it reports an average gain of 4.0 exact-match points across three backbone models. ArXiv · AI/CL/LG's note
The paper targets LLM agents working through task streams where task boundaries and labels are not given. UniMem uses learnable routing tokens to decide whether to rely on retrieval from an episodic buffer or consolidate reliable patterns into parametric memory blocks. The authors say this avoids fixed parameter budgets and uncontrolled growth while reducing dependence on inference-time retrieval. In long-horizon streaming experiments, it reports an average gain of 4.0 exact-match points across three backbone models. ArXiv · AI/CL/LG's note
score 4