Memory fragmentation occurs when non-continuous areas are allocated, requiring the adoption of memory defragmentation policies. Likewise, in NoC-based many-core systems, processing elements (PEs) that communicate with each other may be located in non-contiguous areas, characterizing PE fragmentation. The literature approaching defragmentation often acts on fragmented regions instead of fragmented applications, generating unnecessary task migrations. We propose a reactive and fine-grain defragmentation method. Two sets of experiments demonstrate the effectiveness of the proposal. The first evaluates the defragmentation in an 8x8 system with all PEs executing tasks. The communication cost starts at 2.11 (average hop distance), reaching 1.13 after defragmentation. The second one evaluates the execution time of applications in an actual many-core, showing that fragmentation penalizes the execution time of applications. By applying the defragmentation heuristic, the execution time overhead reduces from 10.6% to 4.2% for an AES benchmark, considering the execution time of task migrations.