Engram memory transfers across LLMs with adapted reader pipeline
Frozen memory tables transfer between AI models—but the reader holds the key.
A new arXiv paper by Mingyuan Li, Guangsheng Yu, Xu Wang, and Shaoxiong Ji tackles a central challenge in large language models: how to reuse learned knowledge across different model backbones. They explore 'Engram-style hashed memory,' a middle ground between retrieval-augmented generation (RAG) and parametric fine-tuning. Instead of storing knowledge in weights or running costly retrievers, Engram maintains an external, addressable memory table consumed by a small learned reader. The question: when moving that frozen memory to a new model, does the memory content or the target-side reader matter more?
Through controlled QA experiments, the researchers found that the reader is the linchpin. A dual-layer, four-branch reader achieved an average score of 38.8, nearly eliminating the performance gap between same-model and cross-model memory reuse. Even without target-side training, a directly compatible provider reader yields substantial utility from the frozen artifact—but adapting the reader pushes performance higher. These findings suggest memory tables can act as reusable, auditable external knowledge artifacts, decoupled from model weights and portable across architectures, with practical implications for modular AI systems and continual learning.
- Engram-style hashed memory stores learned knowledge externally, enabling cross-model transfer without retraining the backbone LLM.
- A dual-layer, four-branch reader achieves a 38.8 average QA score, nearly matching same-model reuse performance.
- Directly compatible reader interfaces allow frozen memory reuse without target-side training; reader adaptation adds further gains.
Why It Matters
External memory artifacts could become portable across LLMs, cutting re-training costs and enabling modular, auditable knowledge updates.