Turn discarded hidden states into an agent memory substrate: MemoryAgentBench accuracy jumps 29.54% to 40.00% with no context reprocessing

arXiv 2607.29032·medium signalXBlueskyLinkedInCopy link

TransMem is a lightweight parametric module that converts sparse historical hidden states into reusable memory representations, using a gating network to intervene dynamically on current hidden states — all at inference time, without reprocessing prior context. Training uses evidence-conditioned self-distillation, where a memory-augmented student sees full context while matching an evidence-only teacher sharing the same frozen backbone. Reported gains are large across three benchmarks: 11.58–29.25 F1 points on LoCoMo, 10.20–13.03 F1 on HotpotQA, and 29.54% to 40.00% accuracy on MemoryAgentBench.