Co-LMLM: Continuous-Query Limited Memory Language Models

Name: Co-LMLM: Continuous-Query Limited Memory Language Models
Uploaded: 2026-10-06T08:59:02.727812+00:00
Duration: 1072 s
Description: CO-LMLM introduces a Limited Memory Language Model that externalizes knowledge into a searchable database instead of memorizing it. It uses continuous keys and textual knowledge values to allow flexible, low-cost vector queries during generation. This enables human-readable, attributable knowledge retrieval while maintaining the controllability and editability benefits of LMLMs.

CO-LMLM introduces a Limited Memory Language Model that externalizes knowledge into a searchable database instead of memorizing it. It uses continuous keys and textual knowledge values to allow flexible, low-cost vector queries during generation. This enables human-readable, attributable knowledge retrieval while maintaining the controllability and editability benefits of LMLMs.

Open the full episode