How ChatGPT's New Memory Feature Works: Technical Analysis and Implementation Details

OpenAI recently launched an additional memory capability called "Chat History" that lets ChatGPT reference past conversations for more personalized interactions. The feature is disabled by default and can be enabled via Settings → Personalization → Reference Chat History.

How the Memory System Works

According to official information, there are two known memory modes: Reference Saved Memory and Reference Chat History . The chat history system can be further divided into three sub‑systems: Current Conversation History , Conversation History , and User Insights .

Saved Memory System

This simpler, user‑controllable system stores custom facts such as a user's name, favorite color, or dietary preferences. Users can add facts with prompts like "Remember that I …" and manage them through the UI. Implementation steps include:

Use a bio tool to approximate the storage mechanism.

Define an LLM call that accepts user messages and an existing fact list, returning an updated fact list or a rejection.

Optionally build a simple UI for viewing and deleting stored facts.

Chat History System

The new chat history feature is more complex and likely contributes to faster assistant responses. It consists of:

Current Conversation History : a short‑term record of recent messages (e.g., within the last day).

Conversation History : broader context from earlier dialogues, providing a concise background.

User Insights : higher‑level, inferred information derived from multiple conversations, such as a user's expertise in Rust asynchronous programming.

Technical implementation may involve multiple vector stores:

Configure two vector spaces named message-content and conversation-summary .

Insert incoming messages into the message-content space; after a period of inactivity, move relevant user data to the conversation-summary space and maintain a third space for indexed summaries.

Additional steps for User Insights could include a weekly Lambda function that queries recent messages, runs an insightUpdate Lambda for each user, and performs clustering to generate concise insights, which are then stored and attached to the model context.

First Wave of Memory Feature Feedback

Early user and expert reactions are polarized. Positive aspects include improved personalization, reduced token usage, and better handling of user intent thanks to the three sub‑systems. However, many users report that the feature often "doesn't work," citing bugs such as a 64‑word storage limit, persistent hallucinations, and other reliability issues.

The article concludes by inviting readers to share their opinions on the ChatGPT memory feature.