Teknosys learning library

8. State tracking, backtracking, and backpropagation

State tracking records relevant conversation turns, task progress, and tool results. A transformer's key/value cache accelerates inference for preceding tokens; it is not durable user memory or a weight update. Applications control persistence, context selection, and privacy.

Backtracking is a search technique: save a state, try a choice, and return to the saved state if a constraint fails. Beam search keeps multiple candidate sequences, but ordinary token generation does not automatically undo earlier text.

Backpropagation computes loss gradients through the network using the chain rule. It is how training assigns responsibility for errors, not how a search algorithm reverses choices.

Educational material. Verify examples against current documentation and applicable licenses. Report a correction.