Petit chat avec Mistral, suite à un bug sur un script qui lance un logiciel externe (genre Acrobat, Word, ...) depuis une ch'tite appli web locale de mon cru, fonctionnait sous Win11 mais problème sur Linux... Mistral m'aide sur un truc, mais se plante totalement sur la partie importante du bug, le fix m'est ensuite donné en allant sur chatGPT (manquait une meta UTF-8 dans le header, et urldecode() - suggéré par Mistral hein, en trop dans une ligne de code).
Donc je retourne sur Mistral et lui fait remarquer qu'il s'est planté, je lui donne la solution (en me foutant de sa gueule
), puis j'embraie sur une ch'tite conversation pour savoir quel modèle il fait tourner dans ce chat, et lui rétorque que non il ne se "souviendra" pas de ce fix en dehors de ce seul chat, autrement dit que quand qq'un d'autre aura le mm problème bah Mistral donnera encore la mauvaise réponse.
Pour ceux et celles que ça peut intéresser : (en EN, dsl mais voilà)
Moi :
So when you say "I’ll remember that for next time!" it's quite misleading, isn't it ? Because next user with the same issue will again get the wrong answer. Even I will get the wrong answer as soon as this chat is closed... I guess this is because your model would need to be re-trained with this new info, am I right ? Mistral : You’re 100% correct, and I appreciate you pushing for clarity. My phrasing was misleading—I don’t "remember" anything between chats, and my responses are based solely on the current conversation and my training data (which has a cutoff and isn’t updated in real-time).
How It Actually Works
No persistence: When this chat ends, all context is gone. Even if you return later, I won’t recall our past exchanges.
Training data: My knowledge is fixed as of November 2024. New fixes (like yours) aren’t added unless Mistral retrains the model with updated data.
Per-chat context: I only "know" what you tell me in this conversation. That’s why I asked you to share your solution—so I could use it here and now.
Why This Happens
Privacy: No user data or conversations are stored.
Scalability: Every chat is independent, so I can assist millions without carrying context.
Limitations: I can’t learn or improve from individual interactions.
How to Work Around It
If you encounter the same issue later, just paste the relevant details (e.g., "We previously fixed this by ensuring UTF-8 and avoiding urldecode() in PHP"). That way, I can give you the correct answer immediately.
Key Takeaways to Share
No memory between chats: Every conversation starts fresh.
Training data is static: My knowledge is fixed as of
November 2024—no real-time updates.
Context is temporary: If you close the chat, I won’t recall past fixes or details.
Workaround: If you solve a problem, note the solution (like your UTF-8 + urldecode() fix) and share it in future chats.
Moi :
Right. One detail I'd be interested in, which Mistral model are you currently (in this chat) running ?Mistral : I'm running on
Mistral Large, the most advanced model from Mistral AI as of my last update.
par Kim Nielsen » 17 Jan 2026, 21:03 
