This sounds incredibly interesting. While I'm not doing anything remotely related to DB at the moment, come June I will find myself with boatloads of free time until I decide to enroll in some place of higher education, and I'm planning to finally give a crack at SF2.2 (the easier route though. I'm not even going to try to finish the far more complicated behaviors), so I may lend some sort of assistance here.
And I'll start with conceptual advice, one of my best skills.
A thought on the 'brain' cells. In a brain, there are processing centers that carry out the calculations and logical decisions dictated to it. Then, there are cells which serve strictly as memory. A 'brain' would require both, say, a veritable 'frontal lobe' and memory cortex. The frontal lobe would perform the actual 'thinking' of the bot, while the memory cortex would simply serve as a memory bus, storing jack in its memvars except what its told to. When the memory cortex grows large enough, it may be of benefit to design a 'memory bus', a middleman between the frontal lobe and the memory that could possibly help reduce processing time (in DB cycles, of course) by dynamically creating and allocating memory arrays for different kinds of data. Of course, storing different 'memories' would require an entirely new metasystem in DNA assembly to equal something of a filesystem and filetypes, but that is another discussion entirely. Read and write times would be measured in multiple cycles, so the memory cortex would be used more like a hard drive than RAM, or, more akin to microprocessors, RAM as opposed to registers. It would be wise to devote different cells of the brain to memory read/write tasks, reading and writing 'memories' to and from storage as required, such as part of the brain deciding to pass the time by performing a statistical analysis of past veggie growth and colony size.
Persistent memory storage alone, as well as any kind of statistical system, not to mention the concept of 'memories' (idea- try and implement some kind of system as used in Dwarf Fortress, where significant events, such as mass deaths, attacks, veggie surplus, etc, are stored as 'memories'.) absolutely REQUIRES the use of auxiliary cells as memory cells. 1000 integer slots is far from enough to store this kind of data when there are other things to be done. If anyone would care to try, it would be nice to see a proof-of-concept bot that would create and maintain "dead-weight" memory cells that it ties to and stores and retrieves data, such as veggie positions and densities over time, its health over time, what enemies it sees, how easy/hard they are to kill, ect. This could lead the way to a bot that can truly learn about, and react to, its surroundings, rather than simply simulating the effect of it through simple forced mutation and adaptation a-la Fruitflies. It would also be good to do so in order to lay the groundwork for memory filetypes and a crude filesystem to be used in later bots. Some kind of interpreter working through a higher level language (I.E., Pybot) would be HIGHLY recommended. Even more so, working from conceptual ideas down through rough logic flowcharts through pseudocode and finally down to DB-DNA is highly recommended as well, if nothing more than to preserve order in the code itself and make sure it stays highly modular.