Claude Memory: A Different Philosophy
How Claude memory works, how it differs from ChatGPT, and what these approaches reveal.
Earlier this week, I dissected ChatGPT's memory system. Since then, I've been doing the same for Claude and realized something remarkable: these two leading AI assistants have built completely opposite memory systems.
In this post, I'll start by breaking down exactly how Claude's memory works—what it stores and how it retrieves information. Then we'll get to the interesting stuff. Why these architectures diverge so dramatically, what that tells us about who uses each assistant and the philosophies driving each product's development, and just how vast the AI memory design space really is.
How it works
Claude's memory system has two fundamental characteristics. First, it starts every conversation with a blank slate, without any preloaded user profiles or conversation history. Memory only activates when you explicitly invoke it. Second, Claude recalls by only referring to your raw conversation history. There are no AI-generated summaries or compressed profiles—just real-time searches through your actual past chats.
When Claude detects memory invocation through phrases like "what did we discuss about," "continue where we left off," or "remember when we talked about," it deploys two retrieval tools that work like web search or code execution—you see them activate in real-time and wait while Claude searches through your history. Once the search completes, Claude synthesizes the retrieved conversations to answer your question or continue the discussion.
Conversation Search
The conversation_search tool helps with keyword and topic-based searches across your entire conversation history. When I asked "Hey, can you recall our past conversations about Chandni Chowk?" (a historic neighborhood in Delhi), Claude found 9 relevant conversations—from when I explored its founding by Princess Jahanara Begum in 1650 to my queries about the best galouti kebabs at Karim's and stuffed parathas at Paranthe Wali Gali. Claude synthesized these scattered discussions into a coherent summary of my Chandni Chowk explorations.
When you ask about multiple topics, Claude runs separate searches sequentially. In my past job as a crypto researcher, I used Claude extensively as an editor. When I asked "Tell me all the conversations we've had about either Michelangelo or Chainflip or Solana," Claude ran three separate searches—one for my Michelangelo analogies for neural networks, another for Chainflip's cross-chain protocol work, and a third for Solana's technical architecture. It found 22 conversations across these searches and delivered a unified response with direct links to each chat.
... continue reading