XNews

Reverse engineering finds ChatGPT uses a four-tier context memory, not a single vector DB/RAG

Reverse engineering finds ChatGPT uses a four-tier context memory, not a single vector DB/RAG

An analysis of ChatGPT’s memory claims it relies on a four-layer context structure—session metadata, long-term user memory (key facts), recent conversation summaries (~15 conversations), and the current session sliding window—rather than exhaustive vector DB/RAG; this design balances personalization with latency and cost and provides a 'pseudo-omniscient' user experience via core facts plus recent summaries.

Source posts