What is delta-mem?

delta-Mem is a research project introducing a new memory system for large language models, the AI systems that power chatbots and similar tools. The problem it addresses is that current models either keep growing their context window, which is expensive to process, or rely on static knowledge baked in during training. delta-Mem proposes a compact, updatable memory that the model can read and write during a conversation, without ballooning compute costs. The core mechanism is called delta-rule learning. As each new piece of information arrives, the model projects it into a smaller memory space and writes an update using a mathematical rule. This lets the model retain earlier conversation history without reprocessing everything each turn. The memory sits on top of an existing language model as a lightweight adapter, it does not replace the base model. Three write strategies called TSW, SSW, and MSW are included, differing in how memory updates get applied. The repository contains the delta-Mem implementation, training scripts, evaluation benchmarks, and an interactive chat demo. It is written in Python and designed to run on NVIDIA GPUs. A pretrained adapter is available for download and can be loaded with a short code snippet. The full README is longer than what was shown. This project is for researchers and engineers working on long context AI agents, memory augmented language models, or efficient inference for conversational systems. The evaluation suite in the repository covers several benchmarks, including LoCoMo, HotpotQA, IFEval, and GPQA, giving a broad picture of how the memory mechanism affects reasoning and retrieval quality across different task types.