A complementary learning system for continual episodic memory in large language models
Humans retain memories of individual experiences for a lifetime, an ability attributed to a complementary learning system in which a fast process encodes episodes and a slow process integrates them into semantic knowledge. In classical Hebbian models such as Hopfield networks, memory traces are superposed in shared weights. This makes learning naturally continual but causes strong interference…
Humans possess the remarkable ability to retain memories of individual experiences throughout their lives, a feat achieved through a complementary learning system. This system operates in two stages: a rapid process encodes episodes, while a slower process integrates them into long-term semantic knowledge. In traditional learning models like Hopfield networks, memories are superimposed onto shared weights, leading to interference and a phenomenon known as catastrophic forgetting in deep neural networks.
To better understand and replicate this human memory process, researchers have employed a large language model as a testbed for modeling continual episodic memory. The pretrained weights of the language model serve as the semantic context within which new experiences are embedded. A key component of this model is a hippocampus-like module responsible for assigning each new episode to a unique, extremely sparse low-rank adapter.
This adapter ensures that each memory remains distinct and unencumbered by interference from other memories.
During recall, competitive gating mechanisms select among these separated traces, enabling the model to retrieve the relevant information efficiently. The model's ability to handle up to 1,000 factual and autobiographical episodes is remarkable, with each adapter requiring only 2-3 parameters per token while maintaining excellent recall rates.
This efficient storage system is complemented by a retrieval-augmented generation mechanism that reconstructs the selected episode within the context of the model, allowing for high-accuracy question answering over the stored memories.
To simulate the slow cortical consolidation process responsible for integrating memories into long-term storage, the model incorporates batch replay of the base weights. This consolidation phase fine-tunes the original pretrained weights, enabling the model to reconstruct and directly answer questions based on stored episodic memories without the need for individual adapters.
By integrating both fast storage and slow consolidation mechanisms within a single language model, researchers have successfully created a neural-network model that accurately emulates the key functional features of human memory, demonstrating the potential for developing artificial systems capable of replicating the remarkable capabilities of the human brain.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
