This paper addresses the inefficiency of storing and processing long conversational histories and documents by introducing a machine-facing compression interface that skips human-readable reconstruction entirely. LatentPress converts text directly into continuous soft tokens that a frozen language model consumes through its input-embedding layer, training only a small adapter representing roughly 0.1% of decoder parameters to perform the mapping. The approach achieves 4-16x compression while matching or exceeding uncompressed baseline accuracy on memory QA and document understanding tasks. It also delivers practical efficiency gains, encoding in 43 milliseconds per conversation and running inference 5-9x faster than processing raw context.
