Reverse-engineering ChatGPT's memory: 4 layers, no vector DB, no RAG

HowDevelop · x · 2026-09-02

Developer Manthan Gupta reverse-engineered ChatGPT's memory system through conversation experiments and published it as the first of a 7-part blog series on agentic memory, covering ChatGPT, Claude, Hermes, OpenClaw, and voice-agent memory designs.

Key finding: ChatGPT's memory is far simpler than expected — no vector databases, no RAG over conversation history. Its context has six slots: system instructions, developer instructions, ephemeral session metadata, long-term user memory, summaries of recent chats, and the current session window.

Memory splits into four layers: ephemeral session metadata (device, browser, location, subscription tier), explicit long-term facts (the author found 33 stored facts about himself), lightweight summaries of past conversations, and a sliding window of the current chat. He argues this layered direct-injection approach may beat traditional retrieval systems. All findings come from behavioral reverse-engineering; OpenAI hasn't published these details.

Original post →

More from coding & agent

coding & agent channel →