Partition agent memory into independent heads — retention jumps from under 30% to 74% at 896K tokens, training-free

arXiv 2607.01523·high signal

arXiv 2607.01523 (July 1, Li/Yeh/Li, UW-Madison) diagnoses recurrent memory agents precisely: the bottleneck isn't capturing information, it's *retaining* it — retention falls below 30% at 896K tokens because every consolidation step overwrites one monolithic memory block. Multi-Head Recurrent Memory (MHM) partitions memory into independent heads with a select-then-update strategy so each write touches one head and leaves the rest protected; the MHM-LRU variant picks the least-recently-updated head, guaranteeing uniform utilization at zero extra token cost. Results: 73.96% retention and 49.74% accuracy on RULER-HQA at 896K (vs 21.62% for MemAgent), and 41.41% vs 25.26% on BABILong at 1M. It's training-free and generalizes across model families, so it's a scratchpad-partitioning change you can make this week in any compaction loop you already run.