Large Language Models (LLMs) face challenges in acquiring new knowledge post-training due to high costs, slow updates, or input size limits. Researchers developed MeMo—a modular memory model separate from the main LLM—that encodes new data efficiently without retraining the entire system. MeMo combines a small MEMORY model, trained on distilled QA pairs of updated knowledge, with a frozen EXECUTIVE reasoning LLM that queries this MEMORY as an oracle. This design avoids drawbacks of traditional methods like retrieval-augmented generation (RAG), fine-tuning issues like catastrophic forgetting, and latent memory constraints. By updating only the MEMORY model and merging with new knowledge models, MeMo supports continuous, cost-effective updates while protecting reasoning capabilities. Testing across benchmarks showed MeMo surpasses existing retrieval systems, handling noisy data robustly and enabling instant performance boosts by switching EXECUTIVE models. However, initial training for reflections is resource-intensive and MEMORY capacity limits exist. MeMo suits enterprises needing complex, multi-document reasoning and offers an effective hybrid approach alongside traditional retrieval.
Back
