The challenge of segmenting online handwritten Tamil words has hardly been investigated. In this paper, we report a neuroscience-inspired, lexicon-free approach to segment Tamil words into its constituent symbols (recognizable entities). Based on a simple dominant overlap criterion, the word is grossly segmented into candidate symbols (stroke groups). However, this segmentation is not fully reliable because of varying writing styles resulting in varying levels of overlap. Taking cues from vertebrate visual perception, we utilize both feature based attention and feedback from the classifier to detect possible wrong segmentations. This attention-feedback segmentation (AFS) strategy splits or merges the stroke groups to correct the segmentation errors and forms valid symbols. This maiden attempt on segmentation is tested on 10000 handwritten words collected from hundreds of writers. The efficacy of AFS in segmentation and improving the recognition performance of the handwriting system is amply demonstrated. Our results show a segmentation accuracy of over 99% at symbol level.