MemFold: 256-vector soft memory doubles context and speeds up inference

Researchers introduced MemFold, a method that compresses arbitrarily long dialogue history into 256 continuous latent vectors the model reads like ordinary tokens. It scales from 32K to 128K tokens of raw history without increasing input footprint while matching or beating baselines on PersonaMem-128K, PrefEval and LongMemEval.

• Memory is fixed at 256 vectors regardless of history length
• Context scales from 32K to 128K tokens without growing input
• Method achieves highest accuracy on every benchmark and backbone
• Average end-to-end inference cost is lower across three backbones