Memini: Multi-Timescale Memory for Continual LLM Knowledge Updating

A new paper on arXiv proposes Memini, an external memory system for large language models (LLMs) that mimics biological memory dynamics. Unlike traditional static or explicitly managed memory, Memini uses a directed graph where each edge has two coupled internal variables—one fast, one slow—following the Benna-Fusi model of synaptic consolidation. This design allows episodic sensitivity, gradual consolidation, and selective forgetting to emerge from a single mechanism, enabling LLMs to continuously update knowledge without retraining. The paper argues that external memory should function as a learning substrate that reorganizes through its own dynamics, rather than being a static repository. The work is categorized under Computer Science > Machine Learning and was submitted to arXiv on May 9, 2025.