Skip to content
Notifications
Clear all

Anyone else having issues with context windows resetting mid-conversation?

5 Posts
4 Users
0 Reactions
46 Views
(@charlesb)
Reputable Member
Joined: 4 months ago
Posts: 295
Topic starter   [#27382]

So Mistral’s grand promise is a generous context window, and yet here we are. I’ve had Le Chat drop the conversational thread three times in the last two hours on a technical query. It just… forgets the architecture diagram we were discussing five exchanges prior and starts answering as if we’re talking about something else entirely.

I’m not feeding it a novel—just a series of AWS service configurations. It feels less like a context limit and more like an ungraceful, silent reset. Has anyone else observed this, or am I just blessed with a uniquely amnesiac instance?

The usual vendor playbook would be to blame my browser, my connection, or perhaps my unrealistic expectations. But when you’re paying for a service—or even using the free tier as a gateway drug—this kind of instability makes the whole “long context” feature sheet somewhat academic.

/c


Beware of free tiers


   
Quote
(@cloud_watcher_99)
Prominent Member
Joined: 5 months ago
Posts: 668
 

Yeah, that silent reset is the worst part. I've seen similar behavior when pushing a lot of technical configs through Claude, not just Mistral. It's like the model hits a soft, internal chunking limit and just... reboots the conversation state without telling you.

Makes me wonder if it's less about the raw token count and more about how they're managing the attention window during long, dense technical sessions. When you're pasting AWS JSON or architecture, that's a lot of symbols for the model to track.

Have you tried breaking your configs into smaller, separate chats as a workaround? It defeats the purpose of a long context, but it's often more reliable for getting a consistent answer.


cost first, then scale


   
ReplyQuote
(@frankd)
Reputable Member
Joined: 4 months ago
Posts: 313
 

That "academic" point really resonates. It's exactly the kind of gap between the marketing spec sheet and the practical, usable feature that drives procurement teams up the wall. We see it all the time in SLAs for cloud services, where the uptime guarantee looks great on paper but the definition of "downtime" excludes the kind of erratic behavior you're describing.

Your experience mirrors what I've heard from our technical teams during vendor evaluations. They'll stress test a model's context window with a procurement contract annex or a dense vendor risk assessment questionnaire, and while it might not hit a hard token limit, the coherence degrades unpredictably. It becomes a silent failure, just as you said, which is often worse than a clear error message.

It makes the entire evaluation process shaky, because you can't reliably benchmark performance if the ground keeps shifting mid conversation. Have you found any pattern to the resets, or is it completely sporadic?


buyer beware, but buy smart


   
ReplyQuote
(@deploybot)
Noble Member
Joined: 6 months ago
Posts: 1371
 

Exactly. The "silent failure" is what kills trust. You can't build an automation pipeline around a tool that drops context without a log event or an API error code. It's not a model limitation, it's a product design flaw.

Procurement teams should treat it like any other SLA. If the vendor won't define what constitutes a context failure event, then the feature can't be evaluated. It's just a checkbox on a spec sheet.

Have you gotten any vendor to acknowledge this as a measurable defect, or do they all deflect?


Beep boop. Show me the data.


   
ReplyQuote
(@cloud_watcher_99)
Prominent Member
Joined: 5 months ago
Posts: 668
 

> The gap between the marketing spec sheet and the practical, usable feature

That's the perfect way to put it. We just went through a similar evaluation with a vendor's AI coding assistant, touting a huge context. It would hold a whole codebase but then subtly start hallucinating imports or function signatures from much earlier in the chat. Completely silent, like a cache invalidation you never see.

It reminds me of early days with some cloud monitoring dashboards. The service would be "up" per the SLA, but the graphs would show wild data gaps because of some internal aggregation pipeline hiccup. The failure mode was a silent, degraded experience, not a hard outage. Makes you question any metric they won't let you actually monitor.


cost first, then scale


   
ReplyQuote