🤖 AI Summary
This work addresses the limitation of existing long-running language agents in selectively internalizing experiences, as they struggle to distinguish high-value, persistently useful knowledge from retrievable factual memories. To overcome this, the authors propose the EVAF mechanism, which integrates an Echo-Valence Attractor Field with gated LoRA to enable parameter-level experience consolidation guided by value and surprise signals. A test–retest protocol is introduced to evaluate behavioral persistence under interference. By decoupling memory access from memory depth, the system simultaneously retains retrievable facts and internalizes critical experiential knowledge. Experiments on GPT-2 and TinyLlama demonstrate that EVAF significantly enhances behavioral stability compared to baselines—including frozen models, pure retrieval, and ungated continual updating—while maintaining low parameter drift and minimal cross-character contamination.
📝 Abstract
Long-running language agents need mechanisms for deciding which experiences should persist after the working context is gone. Retrieval systems can reinsert past text, but they do not by themselves show that an experience has been selectively consolidated into the model's own behavior. We introduce EVAF, an Echo-Valence Attractor Field mechanism for gated LoRA consolidation, and a test-retest protocol for measuring selective parametric consolidation under controlled interference. Across GPT-2 and TinyLlama, EVAF preferentially consolidates high-valence, high-surprise experiences while preserving retrieval-accessible factual memory through a complementary routed memory path. Test-retest measurements show stronger post-interference behavioral persistence than frozen, retrieval-only, and ungated continual-update baselines, while keeping parameter drift and cross-persona contamination low. The results support a separation between memory access and memory depth: retrieving a fact and internalizing an experience are distinct computational operations.