The ability to detect and use environmental patterns is critical for survival and central to many cognitive functions, including those unique to humans. We build cognitive maps from regularities in object positions to navigate our surroundings (e.g., chairs near tables) and segment continuous speech into words based on recurring phoneme patterns. This statistical learning often happens implicitly - without intention or clear awareness of the learned patterns - and is preserved across species, developmental stages and states of consciousness (e.g., coma, wakefulness). However, it remains unknown whether conscious awareness of inputs is necessary for statistical learning. Can patterns presented below the threshold of conscious perception still be learned? Here, we built on recent evidence for non-conscious working memory to test whether non-conscious perception supports sequence learning. Combining behavioral and eye-tracking measures, we found clear signatures for non-conscious learning of spatial sequences: After exposure to fixed pairs of target locations, human observers accurately guessed the second target when cued with the first - even when they reported not having seen the sequence. This above-chance localization was accompanied by anticipatory eye-movement signals toward the second target, suggesting non-conscious prediction mechanisms. These findings extend our understanding of non-conscious working memory and demonstrate that statistical learning may occur even in the absence of conscious perception, highlighting the broader role non-conscious processes play in shaping cognition.