What is an LMM (Large Memory Model)?

LMM stands for Large Memory Model. It’s a newer kind of AI architecture that adds a dedicated memory module to the standard transformer (the design behind ChatGPT and Claude), so the model can store, retrieve, and selectively forget information across very long contexts.

Standard large language models read everything in their current context window and then drop it as soon as the session ends. They have no persistent place to keep things. An LMM bolts on an explicit memory bank that the model can write to, read from, and selectively update, much like a notebook the model carries with it.

What’s actually different about it

The reference architecture is a 2025 paper called LM2 (Large Memory Model). It’s a decoder-only transformer enhanced with an auxiliary memory module that interacts with input tokens via cross-attention. The memory updates through gating mechanisms — forget gates and input gates borrowed from the older LSTM family of networks — which decide what to keep and what to drop.

On long-context reasoning benchmarks (BABILong), LM2 reportedly outperforms a memory-augmented RMT model by 37.1% and a baseline Llama-3.2 by 86.3% on average. Those are large gaps, but only on tasks that specifically test multi-step reasoning across long contexts — the cases standard transformers struggle with.

Why it matters

LLMs are great at one-shot tasks but stumble when they need to reason across many steps, hold relationships in mind, or carry knowledge from one session to the next. The workarounds today — bigger context windows, retrieval-augmented generation, long-term memory tools — are bolted-on solutions to a structural gap.

LMMs are designed for exactly that gap. Long-term dependencies, multi-step reasoning, and the kind of remembering that current chatbots can only fake by being re-fed their own transcripts.

One caveat about the term

LMM is also occasionally used to describe a broader class of memory-augmented models — for example, Memory-Augmented Large Multimodal Models for video understanding. The acronym hasn’t fully settled. The cleanest single reference is the LM2 paper on arXiv.

Related

• What is an API?
• What is GitHub?