TransMem turns LLM hidden states into reusable memory
New method improves LLM long-context reasoning by 11-29% with lightweight memory
A team of researchers led by Haodong Lei has proposed TransMem, a novel approach to enhancing large language model (LLM) agents' long-context reasoning capabilities. Presented in their arXiv paper (arXiv:2607.29032), TransMem addresses a critical limitation in current LLM systems: the underutilization of previously computed representations during extended interactions.
TransMem introduces a lightweight parametric memory module that transforms sparse historical hidden states from a frozen LLM backbone into reusable memory representations. Unlike traditional methods that require re-encoding entire contexts, TransMem uses a gating network to dynamically apply latent interventions to current hidden states. The researchers employ evidence-conditioned self-distillation, where a memory-augmented student model matches the predictive distribution of an evidence-only teacher with the same frozen backbone. The approach demonstrated consistent improvements across multiple benchmarks, including LoCoMo (11.58-29.25 F1 points), HotpotQA (10.20-13.03 F1 points), and MemoryAgentBench (accuracy increased from 29.54% to 40.00%). The code is available on Hugging Face.
- TransMem transforms sparse hidden states from frozen LLMs into reusable memory representations
- Improves F1 scores by 11.58–29.25% on LoCoMo and 10.20–13.03% on HotpotQA benchmark
- MemoryAgentBench accuracy jumps from 29.54% to 40.00% with the new method
Why It Matters
Enables LLMs to better retain and utilize long-context information without expensive re-encoding, improving real-world agent performance