The authors built a simple memory system for language models that stores not just facts but emotional responses. They identify emotion features in a small Gemma model using sparse autoencoders (tools that decompose model activations into interpretable components), save emotion vectors during an experience, then re-inject part of that emotion signal when the model later recalls similar contexts. Testing on decision tasks, they find that emotion alone improves threat detection but doesn't improve choices; semantic labels (plain facts) alone get 52% correct decisions, but semantic + emotion together hits 80%—replicating the neuroscience finding that emotion amplifies knowledge into action rather than replacing it.
Main takeaways:
- Sparse autoencoders on layer 22 of Gemma 3 1B picked out 310 emotion-specific features with psychologically meaningful structure
- Re-injecting emotion vectors at recall (triggered by context similarity at layer 7) steepens the model's threat-safety gradient compared to no memory
- Emotion echo alone doesn't help decision accuracy; semantic memory alone gets 52% good choices; combining both pushes accuracy to 80%
- The result mirrors Damasio's somatic marker hypothesis: emotion marks help only when paired with factual knowledge
- Demonstrates a concrete technique for conditioning model behavior on past emotional context, not just facts