Yandori Research develops neural technology for compressing continuous information streams into compact representations that preserve meaningful temporal patterns. Its featured system, Yandori, processes time-based batches of articles and combines them into fixed-size “world state” embeddings. Rather than simply reconstructing text, the model learns to identify information that helps predict how those states will change over the next minute, hour, and day.

The system brings together a transformer-based encoder, importance-weighted aggregation, temporal attention, and hierarchical memory buffers. Its self-supervised learning approach uses prediction signals to determine which articles matter, without requiring manual annotation. Memory operates across multiple timescales while maintaining a fixed footprint, making the architecture suited to long-running information streams.

For production use, Yandori describes a minute-by-minute pipeline that discovers and reads RSS and Atom feeds, encodes newly collected articles, and updates its temporal memory. The research also details crawler controls, training methods, and implementation choices for stable operation. Yandori’s work focuses on compact, prediction-oriented representations of evolving information, with potential applications wherever streams of content need to be summarized while retaining useful context.