Last updated May 22, 2026.

Researchers report a breakthrough in long-context memory


A team has demonstrated stable recall across million-token contexts, sidestepping the quadratic cost that has limited long-context models.

The approach

Instead of brute-force attention, the method blends retrieval with a learned memory store that compresses older context.

Open questions

Independent replication is still pending, and the gains on real documents may differ from the synthetic benchmarks used so far.

Get the best sent to your inbox, every month

Once monthly, no spam