To Keep or to Forget: Toward Context-Sensitive Memory in Large Language Models
Abstract
While attention mechanisms share formal connections with associative retrieval, they lack core features of biological memory systems, particularly context-sensitive memory formation and recall. Inspired by recent cellular neurobiological evidence on pyramidal two-point neurons (TPNs), we propose a TPN-inspired memory mechanism for large language models (LLMs) that enables context-sensitive memory writing, reading, and updating through triadic modulation loops. In this framework, global contextual signals modulate local feedforward representations to selectively engage memory operations only when local evidence is coherent with the global internal state. Memory encoding and retrieval emerge from apical-amplification regimes, while memory updating is governed by apical-drive dynamics. This context-sensitive memory gating reduces interference between overlapping representations and enables more stable and coherent associative memory formation and retrieval during ongoing feedforward (FF) processing.