OpenAI Models Exhibit Context Tracking Issue, Answering Older Prompts in Extended Conversations
Key Takeaways
- ▸OpenAI models exhibit context tracking failures in extended conversations, responding to earlier prompts instead of current ones across multiple GPT-5.x variants
- ▸The issue appears to be an underlying model behavior rather than a client-side or single-release problem, affecting both Codex and non-Codex models
- ▸The problem undermines user trust by breaking conversational continuity, one of the fundamental expectations users have of language models
Summary
Users of OpenAI's language models, particularly those using Codex and GPT-5.x variants, have reported a concerning failure mode where models respond to earlier prompts instead of current ones during longer conversation sessions. The issue manifests as models suddenly reverting to answering previous topics or questions, sometimes even after correctly handling the new topic initially, suggesting a loss of conversational context rather than a simple misunderstanding. Multiple GitHub issues and user reports indicate this is not a single-release problem or version-specific bug, but rather an underlying model behavior that becomes more pronounced in longer, more demanding sessions with multiple topic transitions. The problem is particularly noticeable in structured coding environments like Codex, where users can clearly identify when a response belongs to an earlier step in the conversation. While no official root cause has been publicly identified and no clean fix has been released, users report that breaking conversations into separate threads or keeping sessions shorter can help mitigate the issue.
- No official root cause analysis or comprehensive fix has been publicly released, though workarounds include shorter sessions and separate conversation threads
Editorial Opinion
This context tracking issue highlights a critical gap between perceived and actual model reliability. While language models are often sophisticated enough to handle complex multi-topic conversations, this failure mode—where they confidently answer the wrong prompt—is particularly insidious because it erodes the foundational trust in conversational coherence. The persistence of this behavior across multiple model versions and families suggests it may be a deeper architectural challenge in how these models manage attention and memory in extended contexts.
