Grok 4.7’s Long-Context Memory Has Severely Degraded Compared to Previous Versions
**Update:** Based on analysis and verification from several different AIs using Grok's own feedback, we now know why version 4.7 is experiencing these "entire section missing" issues. Simply put, Grok 4.7's underlying objective has been completely overhauled to act as a "project manager for code and engineering tasks," making it utterly unfit for long-form roleplay or long-form novel writing.
The reason is that the behind-the-scenes "logical reasoning data" generated by the Grok 4.7 model consumes space at an extraordinary rate. While a visible reply might only be 100 words, Grok's hidden reasoning behind the scenes can easily reach tens of thousands of words. This rapidly exhausts the context window and forcibly squeezes earlier story chapters out of memory. Furthermore, the model has been made extremely rigid—if it cannot locate data within its immediate visible window, it simply asserts that it does not exist, rather than relying on remaining vague impressions to smoothly continue the narrative as older versions did.
As a result, it can no longer retain prior context on its own. Authors are now forced to treat creative writing like engineering documentation, manually saving updated story summaries every few chapters and using prompts to instruct Grok to review these summaries before every single response. For creators accustomed to a natural, fluid writing process, version 4.7 is, for all practical purposes, completely useless for novel writing.
--------------------------------------------------------------------
**Original post:**
As I mentioned before, I use Grok to assist me in writing dark fiction. In earlier versions, if I referenced events that occurred earlier in the story, it might occasionally get minor details wrong, but it still retained the general narrative flow.
However, after updating to 4.7, Grok's context memory seems to have been completely lost. For instance, I had the main character spend a 7-day vacation in Brazil, spanning a total of 28 chapters. Each chapter had an independent title specifying the time, location, and character, such as "Wiraley • Seven Days in Brazil • Day 3 • Day A". I designed it this way because previously, if generated content contained errors regarding past plotlines, as long as the chapter had a distinct title, I could directly reference it in the prompt to instruct Grok to consult that specific chapter before generating text.
With version 4.7, however, referencing titles in plot summaries no longer works. Even when I issue a direct command like "Search for the response titled 'Wiraley • Seven Days in Brazil • Day 3 • Day A'", all I receive in return is:
>"This response is not in the visible history of the current conversation, and there is no title 'Wiraley • Seven Days in Brazil • Day 3 • Day A' in local files. Searched /workspace, /root, and /tmp; found nothing matching 'Seven Days in Brazil' or 'Day 3 • Day'."
Across a total of 30 chapters (including 2 chapters of character internal monologues) and 30 turns of conversation, the content has completely and utterly vanished from Grok’s context memory system.
What is even more egregious: within the very same conversation, I can paste a character's complete backstory and instruct Grok to repeat it back to me. Grok will do so, yet in the very next prompt as I continue the story, the character acts in complete contradiction to that backstory.
I am being entirely serious—despite the numerous controversies Grok has faced in the past, I never considered canceling my subscription. Now, however, I genuinely feel the urge to unsubscribe.
