GradMem: Learning to Write Context into Memory with Test-Time Gradient Descent

Yuri Kuratov, Matvey Kairov, Aydar Bulatov, Ivan Rodkin, Mikhail Burtsev
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:61806-61829, 2026.

Abstract

Many large language model applications require conditioning on long contexts. Transformers typically support this by storing a large per-layer KV-cache of past activations, which incurs substantial memory overhead. A desirable alternative is compressive memory: read a context once, store it in a compact state, and answer many queries from that state. We study this in a context removal setting, where the model must generate an answer without access to the original context at inference time. We introduce GradMem, which writes context into memory via per-sample test-time optimization. Given a context, GradMem performs a few steps of gradient descent on a small set of prefix memory tokens while keeping model weights frozen. GradMem explicitly optimizes a model-level self-supervised context reconstruction loss, resulting in a loss-driven write operation with iterative error correction, unlike forward-only methods. On associative key–value retrieval, GradMem outperforms forward-only memory writers with the same memory size, and additional gradient steps scale capacity much more effectively than repeated forward writes. We further show that GradMem transfers beyond synthetic benchmarks: with pretrained language models, it attains competitive results on natural language tasks including bAbI and SQuAD variants, relying only on information encoded in memory.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-kuratov26a, title = {{G}rad{M}em: Learning to Write Context into Memory with Test-Time Gradient Descent}, author = {Kuratov, Yuri and Kairov, Matvey and Bulatov, Aydar and Rodkin, Ivan and Burtsev, Mikhail}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {61806--61829}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/kuratov26a/kuratov26a.pdf}, url = {https://proceedings.mlr.press/v306/kuratov26a.html}, abstract = {Many large language model applications require conditioning on long contexts. Transformers typically support this by storing a large per-layer KV-cache of past activations, which incurs substantial memory overhead. A desirable alternative is compressive memory: read a context once, store it in a compact state, and answer many queries from that state. We study this in a context removal setting, where the model must generate an answer without access to the original context at inference time. We introduce GradMem, which writes context into memory via per-sample test-time optimization. Given a context, GradMem performs a few steps of gradient descent on a small set of prefix memory tokens while keeping model weights frozen. GradMem explicitly optimizes a model-level self-supervised context reconstruction loss, resulting in a loss-driven write operation with iterative error correction, unlike forward-only methods. On associative key–value retrieval, GradMem outperforms forward-only memory writers with the same memory size, and additional gradient steps scale capacity much more effectively than repeated forward writes. We further show that GradMem transfers beyond synthetic benchmarks: with pretrained language models, it attains competitive results on natural language tasks including bAbI and SQuAD variants, relying only on information encoded in memory.} }
Endnote
%0 Conference Paper %T GradMem: Learning to Write Context into Memory with Test-Time Gradient Descent %A Yuri Kuratov %A Matvey Kairov %A Aydar Bulatov %A Ivan Rodkin %A Mikhail Burtsev %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-kuratov26a %I PMLR %P 61806--61829 %U https://proceedings.mlr.press/v306/kuratov26a.html %V 306 %X Many large language model applications require conditioning on long contexts. Transformers typically support this by storing a large per-layer KV-cache of past activations, which incurs substantial memory overhead. A desirable alternative is compressive memory: read a context once, store it in a compact state, and answer many queries from that state. We study this in a context removal setting, where the model must generate an answer without access to the original context at inference time. We introduce GradMem, which writes context into memory via per-sample test-time optimization. Given a context, GradMem performs a few steps of gradient descent on a small set of prefix memory tokens while keeping model weights frozen. GradMem explicitly optimizes a model-level self-supervised context reconstruction loss, resulting in a loss-driven write operation with iterative error correction, unlike forward-only methods. On associative key–value retrieval, GradMem outperforms forward-only memory writers with the same memory size, and additional gradient steps scale capacity much more effectively than repeated forward writes. We further show that GradMem transfers beyond synthetic benchmarks: with pretrained language models, it attains competitive results on natural language tasks including bAbI and SQuAD variants, relying only on information encoded in memory.
APA
Kuratov, Y., Kairov, M., Bulatov, A., Rodkin, I. & Burtsev, M.. (2026). GradMem: Learning to Write Context into Memory with Test-Time Gradient Descent. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:61806-61829 Available from https://proceedings.mlr.press/v306/kuratov26a.html.

Related Material