UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams
UniMem routes novel experience into retrieval memory and consolidates recurring patterns into expandable parameters, offering a design for agents that learn across unlabelled task streams.
UniMem uses **learnable routing tokens** to choose between an episodic retrieval buffer and expandable parametric memory. Novel or sparse tasks stay external; recurring, reliable execution patterns are gradually consolidated.
For long-running agents, rethink memory as two complementary paths rather than a single retrieval store. Separate task identification from execution so memory can adapt without explicit task boundaries or a fixed parameter budget.
UniMem uses **learnable routing tokens** to choose between an episodic retrieval buffer and expandable parametric memory. Novel or sparse tasks stay external; recurring, reliable execution patterns are gradually consolidated. For long-running agents, rethink memory as two complementary paths rather than a single retrieval store. Separate task identification from execution so memory can adapt without explicit task boundaries or a fixed parameter budget. The reported **4.0-point average exact-match gain** spans three backbone models on long-horizon streaming sequences. The supplied material does not establish deployment cost, consolidation safety, or performance on production coding-agent workloads.
UniMem changes the memory design space from choosing a retrieval store to routing between episodic retrieval and learned consolidation. It complements prior work on memory value, provenance, and shared context by addressing where recurring execution knowledge should live, but leaves their governance concerns unresolved: the supplied evidence does not show that consolidation is safe, inspectable, or economical in production agents.