How does ChatGPT manage context and memory across multiple turns, such as when asking the 10th question in a conversation?

ChatGPT’s core model operates on a fixed token window (about 4,096 tokens for GPT‑3.5 and 8,192 for GPT‑4). During a conversation, the model receives the concatenated prompt that includes the user’s last message, the assistant’s previous reply, and a few earlier turns. When the conversation grows beyond the window, older turns are truncated from the beginning of the prompt, effectively discarding them from the model’s immediate context. To preserve longer‑term memory, OpenAI’s production system uses a retrieval‑augmented approach: it stores key‑value pairs of past interactions in an external vector store and retrieves the most relevant snippets when generating a response. Additionally, the model can be fine‑tuned with a “memory” prompt that reminds it of user preferences or prior facts. In practice, this means that by the 10th turn, the model still has the most recent few turns in its prompt, but earlier context may be summarized or retrieved from the external store. This hybrid strategy balances the token budget with the need for continuity, allowing ChatGPT to remember user goals and facts across a conversation while still operating within its token limits.