There is. The Scipy Cookbook WebSite includes a complete implementation of a kNN algorithm that can be updated incrementally.
Maybe a few lines of background would be helpful for anyone interested but not familiar with the terminology.
A kNN engine is powered by either of two data representations--the pairwise distances between all points in the dataset stored in a multi-dimensional array (a distance matrix), or a kd-tree, which just stores the data points themselves in a multi-dimensional binary tree.
These are only two operations that a kd-tree-based KNN algorithm needs: you create the tree from the dataset (analogous to the training step performed in batch mode in other ML algorithms), and you search the tree to find 'nearest neighbors' (analogous to the testing step).
Online or incremental training in the context of a KNN algorithm (provided it's based on a kd-tree) means to insert nodes to an already-built kd-tree.
Back to the kd-Tree implementation in the SciPy Cookbook: The specific lines of code responsible for node insertion appear after the comment line "insert node in kd-tree" (in fact, all of the code after that comment are directed to node insertion).
Finally, there is a kd-tree implementation in the spatial module of the SciPy library (scipy.spatial module) called KDTree (scipy.spatial.KDTree) but i don't believe it supports node insertion, at least such a function is not in the Docs (i haven't looked at the source).
