Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2607.29032.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

TransMem is a lightweight inference-time parametric memory module that converts sparse historical hidden states from a frozen LLM into reusable memory via a gating network and evidence-conditioned self-distillation, yielding 11–29 F1 gains on LoCoMo and 10–13 F1 on HotpotQA across model scales.

Concrete architecture and benchmark numbers for a memory layer that avoids re-encoding past context, relevant to anyone building long-horizon LLM agents.

Cited by

No citations on record.