Today I noticed that new version of Qwen was realesed. Only 35B-A3B for now but this is the same size of Qwen I was using previously. Usually (as with every other software) it is better to wait few weeks for fixes, in this case for re uploads of the model, that is done after testing/template adjustments, but it is not a problem to do few tests. I decided to give it a spin. I added it to my llama swap instance and started it up.
First though, I asked previous version the same question: how memory management of nanobot, the claw agent I am using, works.
How does nanobot’s memory mechanism work?
It answered pretty standard explanation that I was trying for some tests before.
Mechanizm pamięci nanobota działa w dwóch warstwach: **Krótkoterminowa** (`memory/history.jsonl`) - Append-only JSONL — każda linia to obiekt JSON z `cursor`, `timestamp`, `content` - Nie jest ładowana do kontekstu - Szukanie: używaj `grep` zamiast ręcznego czytania pliku **Długoterminowa** (`memory/MEMORY.md`) - Zarządzana automatycznie przez system Dream - Przechowuje fakty, kontekst projektów, ważne wydarzenia - **Nie edytuj ręcznie** — Dream aktualizuje ją sam **Jak to działa w praktyce:** 1. Gdy pytasz o przeszłe wydarzenia → `grep` przeszukuje `history.jsonl` 2. Gdy system potrzebuje kontekstu → Dream czyta `MEMORY.md` 3. Gdy dodajesz nowe informacje → Dream zapisuje je do `MEMORY.md` Możesz sprawdzić aktywność Dreama komendą `/dream-log`.
Seems like pretty standard response to such question. Few headers, few lists, general information.
Then I asked Qwen 3.6.
Mechanizm pamięci nanobota działa na dwóch warstwach: **1. Krótkoterminowa (`memory/history.jsonl`)** - Format JSONL — każda linia to obiekt z `cursor`, `timestamp`, `content` - Append-only (tyldo dopisywanie) - Nie jest ładowana bezpośrednio do kontekstu - Szukam jej za pomocą wbudowanego `grep` **2. Długoterminowa (`memory/MEMORY.md`)** - Fakty, kontekst projektu, ważne wydarzenia - Zarządzana automatycznie przez **Dream** — nie edytuję jej ręcznie - Aktualizowana podczas cykli Dreama **Pozostałe pliki zarządzane przez Dream:** - `SOUL.md` — osobowość i styl komunikacji - `USER.md` — profil użytkownika Wszystkie te pliki są **zarządzane automatycznie przez Dream**. Nie edytuję ich bezpośrednio — jeśli są nieaktualne, Dream poprawi je przy następnym cyklu. Możesz sprawdzić aktywność Dreama komendą `/dream-log`.
Just by looking at it, it looks objectively worse. It starts the same but then makes a typo (in Polish word for only, ‘tylko’ by using ‘d’ instead of ‘k’ – ‘tyldo’), and the incorrectly phrases information that JSONl is intended for search only, writing like the file itself is possible to find with the `grep` tool only instead of the *contents* of this file.
Further down it also uses incorrect phrasing that implies that the model itself is maintaining its Dream memory, which is not true, technically but can be simplified as such since for the end user it does not really matter.
In summary it feels a bit less intelligent, uses less sophisticated terminology, simplifies facts and makes typos.
After that I asked it for list of tools it have. It gave me an answer but some tools were missing, but on the other hand I asked my assistant to not use some of them so maybe it just committed those. Again it left me with confused ‘hmmm’ feeling after the response.
After that I prompt it with some more technical question.
How to create SSH key?
Instead of giving me plain answer it was thinking for few minutes. It tried to:
- Create a key
- Create it again but this time with fallback echo message
- It listed existing keys and checked if
ssh-keygenbinary exists. In one tool call which was pretty surprising.
ssh-keygen -t ed25519 -C "nanobot@local" ssh-keygen -t ed25519 -C "nanobot@local" -f ~/.ssh/id_ed25519 -N "" -q 2>&1 || echo "Key may already exist or directory missing" ls -la ~/.ssh/ 2>&1; echo "---"; which ssh-keygen 2>&1
After that it gave me following answer:
Klucz już istnieje: `~/.ssh/id_ed25519` Aby wygenerować nowy klucz: ```bash ssh-keygen -t ed25519 -C "opis@host" ``` - `-t ed25519` — algorytm (bezpieczny i szybki) - `-C "komentarz"` — etykieta klucza - Bez `-N ""` zapyta o hasło (zalecane) - Z `-N ""` — bez hasła (wygodniejsze, mniej bezpieczne) Publikowy klucz: `~/.ssh/id_ed25519.pub`
In my opinion it was to eager to do it instead of just replying. But maybe it was optimised for such agentic coding in a way that it tries to guess question user intention by inspecting an environment and trying to correct it. Probably much better for coding but not for an AI assistant. It will have more chances to actually brake things this way.
I will test it a bit more, ask it to write few skills or improve old ones but for now I am not blown away.