The paper introduces Larimar, a novel architecture designed to enhance Large Language Models (LLMs) with a distributed episodic memory. This architecture aims to efficiently and accurately update the knowledge stored in LLMs without the need for computationally expensive re-training or fine-tuning. Larimar's memory allows for dynamic, one-shot updates of knowledge, demonstrating comparable accuracy to competitive baselines on multiple fact editing benchmarks, while achieving up to 10 times faster speed. The paper also explores mechanisms for selective fact forgetting, information leakage prevention, and input context length generalization, showing their effectiveness. Larimar is implemented using a hierarchical memory model, similar to the Kanerva Machine, where memory writes and reads are interpreted as inference in a generative model. The architecture is trained end-to-end on generic data and does not require access to edits during training. The paper provides a detailed explanation of the model's training and inference mechanisms, as well as experimental results demonstrating Larimar's performance in various tasks, including single and sequential fact editing, selective forgetting, and input context length generalization.The paper introduces Larimar, a novel architecture designed to enhance Large Language Models (LLMs) with a distributed episodic memory. This architecture aims to efficiently and accurately update the knowledge stored in LLMs without the need for computationally expensive re-training or fine-tuning. Larimar's memory allows for dynamic, one-shot updates of knowledge, demonstrating comparable accuracy to competitive baselines on multiple fact editing benchmarks, while achieving up to 10 times faster speed. The paper also explores mechanisms for selective fact forgetting, information leakage prevention, and input context length generalization, showing their effectiveness. Larimar is implemented using a hierarchical memory model, similar to the Kanerva Machine, where memory writes and reads are interpreted as inference in a generative model. The architecture is trained end-to-end on generic data and does not require access to edits during training. The paper provides a detailed explanation of the model's training and inference mechanisms, as well as experimental results demonstrating Larimar's performance in various tasks, including single and sequential fact editing, selective forgetting, and input context length generalization.