The Brain not working as expected (or is it me?)

[Snoopy]

First of all, i read the full announcement regarding the brain and THOUGHT i understand it, but seems not...
https://hashcat.net/forum/thread-7903.html

so i started some tests using brain server and client with client-features=3. i used a very basic setup (md5 to crack and a simple dictionary)

first run (one matching pw in dict-1.txt)

Code:

.\hashcat64.exe -z --brain-client-features=3 -a 0 -m 0 .\hash-1.txt .\dict-1.txt
Recovered........: 1/5 (20.00%) Digests, 0/1 (0.00%) Salts
Progress.........: 10/10 (100.00%)
Rejected.........: 0/10 (0.00%)

worked as expected

second run (added one new matching pw to dict -> dict2.txt

Code:

.\hashcat64.exe -z --brain-client-features=3 -a 0 -m 0 .\hash-1.txt .\dict-2.txt
Recovered........: 2/5 (40.00%) Digests, 0/1 (0.00%) Salts
Progress.........: 11/11 (100.00%)
Rejected.........: 10/11 (90.91%)

worked as expected, rejected 10 old already testet pw , just using the new one, recovered one new hash, perfekt

now i simulated some other work, blabla, my dict is evolving to dict-3 and now im getting a new hashlist from somewhere else (hash-2.txt) with some already known pw (potfile to the rescue) but also some new hashes with "old" pw within dict-3.txt

Code:

.\hashcat64.exe -z --brain-client-features=3 -a 0 -m 0 .\hash-2.txt .\dict-3.txt
Recovered........: 3/10 (30.00%) Digests, 0/1 (0.00%) Salts
Progress.........: 12/12 (100.00%)
Rejected.........: 11/12 (91.67%)

not working as i expected, brain rejected 11 pw from dict-3 but these pw where never testet against this new file hash-2.txt

i thought the brain stores a hash of the input file, to track attacks versus that file but seems not?

so how can i achieve, that the brain tracks attacks but also recognize new files the correct way?