MemFold
A soft memory method that compresses long conversation histories into a fixed set of 256 latent vectors. It uses trainable write and read models to enable long-context reasoning without increasing input token consumption, thereby improving inference efficiency.
Timeline (last 90 days)
MemFold compresses arbitrarily long conversations into a small number of latent vectors, which the reader model consumes as ordinary tokens without needing to process the full original transcription text.
MemFold fixedly uses K=256 continuous vectors, so the model's input footprint does not increase regardless of whether the original history is 32K or 128K tokens.
MemFold's soft memory size does not depend on history length and is size-agnostic.
MemFold uses a dedicated writer model to encode the original conversation history into 256 vectors, with the writer model's parameters being end-to-end trainable.
MemFold's memory vectors are continuous and differentiable, allowing gradients to backpropagate through the memory interface.
MemFold achieved the highest accuracy on every benchmark and every backbone model, tying only with the strongest baseline on a single cell of PrefEval.
MemFold's end-to-end average cost is close to full-context inference, and its overall cost is lower across three backbone models.
MemFold achieves higher accuracy and lower overall cost on PersonaMem-128K, and ties with the strongest baseline on PrefEval.
MemFold's primary training scenario is persona-driven personalization tasks, but it can transfer to PrefEval and LongMemEval without target-domain training.
MemFold is trained using on-policy optimization and confidence-gated distillation, the latter of which requires an additional training phase with a frozen teacher model.
2 more timeline events
All Facts (12)
MemFold compresses arbitrarily long conversations into a small number of latent vectors, which the reader model consumes as ordinary tokens without needing to process the full original transcription text.
50%UnverifiedMemFold fixedly uses K=256 continuous vectors, so the model's input footprint does not increase regardless of whether the original history is 32K or 128K tokens.
50%UnverifiedMemFold's soft memory size does not depend on history length and is size-agnostic.
50%UnverifiedMemFold uses a dedicated writer model to encode the original conversation history into 256 vectors, with the writer model's parameters being end-to-end trainable.
50%UnverifiedMemFold's memory vectors are continuous and differentiable, allowing gradients to backpropagate through the memory interface.
50%UnverifiedMemFold achieved the highest accuracy on every benchmark and every backbone model, tying only with the strongest baseline on a single cell of PrefEval.
50%UnverifiedMemFold's end-to-end average cost is close to full-context inference, and its overall cost is lower across three backbone models.
50%UnverifiedMemFold achieves higher accuracy and lower overall cost on PersonaMem-128K, and ties with the strongest baseline on PrefEval.
50%UnverifiedMemFold's primary training scenario is persona-driven personalization tasks, but it can transfer to PrefEval and LongMemEval without target-domain training.
50%UnverifiedMemFold is trained using on-policy optimization and confidence-gated distillation, the latter of which requires an additional training phase with a frozen teacher model.
50%UnverifiedMemFold's soft memory budget is fixed at 256 vectors, and scaling to larger sizes may require redesigning the architecture.
50%UnverifiedThe authors argue that if the results can be reproduced across more domains, the standard practice of naively concatenating user history should be abandoned.
50%