AI agents built on recurrent memory have a forgetting problem — and a new paper proposes a structural fix that doesn't require retraining any models.
Researchers identified that so-called recurrent memory agents — systems that compress long inputs into a fixed memory window as they go — degrade badly as context grows. The culprit isn't capturing new information; it's holding onto old information while absorbing new. Current designs store memory as a single text block, meaning every update risks erasing what was already there. The proposed solution, Multi-Head Recurrent Memory (MHM), splits memory into separate independent partitions called heads. At each step, only one head gets updated while the rest are locked — protecting prior content by design rather than by hoping the model behaves itself. A lightweight version, MHM-LRU, cycles through heads in order of least-recent use at no extra token cost. On a standard long-context benchmark at nearly 900,000 tokens, retention jumped from under 30% to nearly 74%.
The gains matter because the race to extend AI context windows has mostly focused on transformer-based approaches — burning compute on attention across millions of tokens. Recurrent memory is the cheaper alternative, but its reliability problems have limited adoption. A training-free fix that works across model families and scales is exactly the kind of result that could make recurrent agents practical for real deployments.
Of course, benchmark gains and production behavior are different things — 73.96% retention at 896K tokens is a marked improvement, but it still means roughly one in four pieces of information gets lost.