TransMem converts discarded hidden states into agent memory

TransMem takes the hidden states a language model normally discards after each turn and turns them into memory it can reuse later, per the paper posted to arXiv. That matters for agents running long conversations or long documents, where recomputing context from scratch means reprocessing everything on every turn. TransMem skips that step, intervening on the model's current hidden states at inference time instead.

A gating network decides which stored pieces of history to pull back in. Training relies on evidence-conditioned self-distillation. A memory-augmented student model sees the full context and learns to match an evidence-only teacher model on a shared frozen backbone. That setup teaches the student to compress history into something reusable without holding the whole transcript in front of it every time.

The benchmark numbers back it up. On LoCoMo, a long-conversation memory test, TransMem gains 11.58 to 29.25 F1 points. On HotpotQA it adds 10.20 to 13.03 F1 points. On MemoryAgentBench, accuracy climbs from 29.54% to 40.00%.

The catch

A jump from roughly 30% to 40% accuracy is real progress. But it's a reminder that an agent running TransMem still gets memory-dependent tasks wrong more than six times out of ten. The paper doesn't say how the frozen-backbone requirement holds up if a team swaps in a different or fine-tuned model later. That matters for anyone trying to bolt this onto an existing agent stack instead of building around it from scratch.