TransMem: Transforming Hidden States into Memory for Large Language Models

Researchers introduced TransMem, an inference-time memory module for frozen large language models that turns sparse past hidden states into reusable representations.

It applies memory through a gating network and uses evidence-conditioned self-distillation to learn when to rely on memory without adding task-specific knowledge.

The approach offers a lightweight way to reduce repeated context encoding and improve reasoning over long interaction histories.

TransMem delivers gains on LoCoMo, HotpotQA, and MemoryAgentBench across multiple model sizes and architectures.