The TPM is a piece of secure* hardware that provides various cryptographic services to the host system. It is important to understand that it is not a crypto accelerator. It is a place to store keys and small pieces of data (like hashes, counters) where it's difficult for someone to extract them even if they have physical access.
The TPM is designed to support trusted computing, a rather splendid set of extensions to the x86 architecture that let you do remote attestation, software sealing and other things. Or at least it would be splendid if it had been really finished off and pushed to completion by the designers. Unfortunately due to various political issues it exists in a quasi-finished, semi-broken state which only experts can use. Without a doubt you have never run any software in a TC environment.
As part of that role, the TPM provides some permanent storage in the form of NVRAM. Because the TPM is designed to be as cheap as possible, it has a limited number of write cycles. Normally you're meant to store Intel TXT launch control policies and sealed keys there, but Pond uses it in a different way by storing keys there that it encrypts local data with. By erasing the key in the TPM chips memory area, the data on disk is effectively destroyed too.
This is useful because modern "disks" are often SSD drives, or physical metal disks that use log structured file systems. Because flash memory has a limited number of write cycles per cell, internally SSDs have firmware that remap writes from logical addresses to different physical addresses, the goal is to avoid wearing down the drive and extend its useful life. Normally it doesn't matter, but if you want to delete data such that it's really really gone, it obviously poses a problem. Using TPM NVRAM solves it, albiet, at a high usability cost.
*note: actual tamper resistance of real-world TPM chips is not something that seems to have been studied much