If a long AI session starts producing weaker answers, it’s rarely a mistake you made. It’s the model running low on room to think.
Cristiano Winckler, Somebody Digital’s Director of Digital Operations, compares it to how a person’s memory works under a hard limit. “Imagine your brain has a certain capacity. I can only read, I don’t know, one million words per day. And after that, I’m going to need to write my notes in a journal, and I will go back to that journal after those one million words have been consumed. Of course, you’re not going to have all the detail in that journal. You’re going to try and write as much information as possible, but you might miss a few things here and there. That’s kind of how AI works.”
Every AI tool has a limit measured in tokens, the unit used to track how much a model can hold in memory at once. Some tools cap out around 200,000 tokens. Others handle a million. The longer a conversation runs, the more of that capacity gets used, and once it’s close to full, the model starts to forget. “It’s not going to remember the details of certain conversations you have had at the beginning,” Winckler says. “It will forget nuances of certain things, instructions you gave a few moments ago.”
It happens without a warning light
When a tool nears its limit, it compacts its own memory to free up space. That process comes at a cost. “A lot of the context and the detail of the conversation is lost,” Winckler says. Models with smaller context windows compact sooner and more often, and doing it repeatedly compounds the damage. “You can let it compact once or twice,” he says, “but after that, you’re going to start losing a lot of context, and the details will be lost.”
None of this comes with a warning. “That happens silently,” Winckler says. “It doesn’t warn you that it forgot a few things; it doesn’t have all the details anymore. That’s the trick.” The signals are subtle: a response that’s slightly off, a detail that doesn’t match anything discussed, or an answer that feels less sharp than it did an hour earlier.
The fix is a proper handover, instead of a fresh start
Somebody Digital’s answer isn’t to avoid long sessions. It’s to end them deliberately, before the quality drops. “The moment we start flagging that context is degrading, and the quality of the output is no longer good, this is when we stop and say, ‘ Write a handover,” Winckler says. That handover captures the objectives, the work completed so far, and what’s still open, then gets passed to a new session in the same project folder.
Getting the handover right takes judgment. “It has to be thorough; it has to contain a lot of detail,” Winckler says, “but not too much that when the next session reads that handover document, it already exhausts the whole context window.” Written well, it keeps a project moving at a high standard without losing the thread. Written carelessly, it just delays the same problem.
Should leadership set the rules?
John Wilkes puts the question to Winckler directly: Should this be something leadership mandates, rather than leaving individuals to notice the drop in quality on their own? Winckler’s answer is education over policy. Somebody Digital runs a regular internal session built specifically for this. “We have our famous, or infamous, Friday AI sessions, where we share knowledge, where we try to educate our team members with the latest tips, tricks, and best practices,” he says. The guidance itself doesn’t need to be long. “This is not a long document with five to ten bullet points that will do the trick,” he says, “but that alone will increase the quality of how your team members utilize AI significantly.”
For any team leaning on AI for client-facing or accuracy-sensitive work, this is the practical takeaway: Know what context degradation looks like, watch for the subtle signs, and build the handover habit before the quality drop becomes the client’s problem.
Context degradation occurs when an AI model reaches its token limit and begins to “forget” earlier information, instructions, or nuances from the conversation. As the session continues, the model may start providing less accurate or lower-quality responses without alerting the user.
There is no explicit warning. You may notice subtle signs, such as responses becoming slightly off-target, details not matching previous parts of the discussion, or answers that feel less sharp and accurate than they were at the beginning of the session.
The most effective practice is to implement a “handover” habit. When you notice signs that the output quality is dropping, stop the current session, write a thorough summary of the objectives, completed work, and open tasks, and then start a fresh session with that information.
Rather than rigid policies, education is generally more effective. Sharing best practices, tips, and tricks during regular team sessions can significantly improve how team members utilize AI tools, empowering them to identify and handle context degradation on their own.
A well-written handover captures critical information, allowing a project to continue seamlessly in a new session. However, it must be balanced; if the summary is too long, it might immediately exhaust the new session’s context window, essentially delaying the problem rather than solving it.
