Description
When building a multi-step Assistant workflow, the model sometimes forgets earlier instructions or skips important context, even though all messages are provided correctly. This inconsistency makes it difficult to build reliable chained tasks.
Steps to Reproduce
- Start an Assistant thread with clear instructions in the system message.
- Add user turns and Assistant turns to build context (e.g., a form intake followed by a summary).
- Add another user turn that depends on the previous context.
- Observe that the Assistant may produce an irrelevant or incomplete response, ignoring parts of the prior thread.
Expected Result
The Assistant should faithfully use the full conversation history to produce a coherent answer.
Actual Result
- Sometimes forgets previous variables or facts.
- Reverts to generic answers.
- Ignores clarifications given earlier in the same thread.
Impact
Breaks multi-agent pipelines or form workflows that rely on consistent context across turns.
Environment
- Assistant API (v2)
- Model: GPT-4, GPT-4 Turbo, GPT-4o-mini
- Observed June 2025
Additional Context
Occurs more often if the system prompt is large or if many steps are chained in a single thread.
Suggested Priority
High - makes multi-turn agent workflows unreliable.