Long-running AI agents have a memory problem — and it is not just about storage.
Researchers have introduced EVAF, an Echo-Valence Attractor Field mechanism designed to decide which experiences a language model should consolidate directly into its parameters versus leave accessible only through retrieval. Tested on GPT-2 and TinyLlama, EVAF uses gated LoRA updates to preferentially bake in high-valence, high-surprise experiences — the ones that, in theory, most warrant changing the model's actual behavior. A complementary routed memory path keeps ordinary factual recall intact without overwriting it. The team also developed a test-retest protocol to measure whether consolidation actually holds up under interference, a more demanding bar than simply checking whether something can be retrieved.
The results matter because they draw a clean line between two things the field often conflates: memory access and memory depth. Retrieval-augmented systems can reproduce a past fact on demand, but that is not the same as a model having internalized an experience in a way that shapes future behavior. EVAF outperformed frozen, retrieval-only, and ungated continual-update baselines on post-interference behavioral persistence while keeping parameter drift and cross-persona contamination low — two failure modes that have quietly plagued continual learning research.
The caveats are real: GPT-2 and TinyLlama are small models, and whether this mechanism scales to the parameter counts that actually run in production is an open question the paper does not answer. Still, the test-retest framing is the genuinely useful contribution here — giving the field a reproducible way to measure whether an agent has learned something versus merely looked it up.