I keep lookup table with pages. Every contains location of page in memory, in vram and dirty flag. If a texture (or vram area during move to frame buffer, etc.) is being used I upload it into vram. It stays there as long, as possible. If there is a write to the page, it's being marked as dirty, and sent again into vram when it is to be used (if the depth of texture is the same, otherwise it's being freed on vram side and uploaded anew). When there is no more place for textures, the vram manager tries to find the places that are big enough to hold a texture page of said depth, and replaces the least recently used one with the new one. The problem is that:

1. Not everything is implemented properly yet.
2. It isn't always as efficient as I thought it would be.
3. Everything gets screwed more often than not.
4. Quite often the vram gets too fragmented, which makes the manager free some used textures (whole texture pages, as they can contain few textures with the same depth but different cluts), while there would be much more than enough place if the vram has been defragmented (it tries to intercept such cases, then frees all vram and just puts all the needed textures anew - which seems to be faster in most of the cases). Yes, I know PS2 has more Vram than PSX, but we need quite a big (two of them ) framebuffer and the manager/lut thing isn't perfect.