Emotion Vector Re-Injection Enhances Language Model Recall

A new arXiv paper introduces a method to give language models episodic-like memory by re-injecting emotion vectors during recall. The authors argue that current models store semantic memory (what happened) but lack the 'how it felt' component, analogous to Damasio's somatic marker hypothesis. Using Gemma 3 1B-IT with Gemma Scope 2 sparse autoencoders, they identified 310 emotion-exclusive features at layer 22. They construct emotion vectors during experience and partially re-inject them at recall, triggered by context similarity at layer 7. Four conditions were tested: no memory, semantic labels only, emotion echo only, and both. The emotion echo alone steepened emotional orientation, paralleling Damasio's framework where emotional markers aid decision-making.