soma is a continual learning architecture. the accompanying paper published by aaai 2026 details the use of a ‘trace bank’ to provide a spectral basis for learning. this enables continuous adaptation at any scale, on any stream of data.

the continual learning architecture affords several key advantages: the training efficiency means it can run on very constrained hardware, checkpoints can be forked and trained further very easily, and the cumulative nature of training means checkpoints get better with experience. logOS is a platform where users can share soma checkpoints. the runtime is free.

checkpoints on logOS can be forked: downloaded, trained further on new data, and re-uploaded. lineage is cryptographic — every checkpoint carries an immutable record of its full ancestry. you can always trace a model back to its roots.

the barrier to entry for ai training has just dropped significantly.