<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>Lessons from Studying Every Memory System — Shlok Khemani, Independent</title>
        <link>https://video.ut0pia.org/videos/watch/9b396638-fc9e-4ef3-8129-f4f7c480ad9d</link>
        <description>A profile ChatGPT keeps on Shlok Khemani says he travelled to Turkey in 2025. He never has. The memory came from conversations where he was choosing between Turkey and Thailand, he went to Thailand, and the profile kept both with overlapping dates. What bothers him is not the mistake but the incuriosity: nothing notices the conflict, and the evidence to settle it was sitting in his email as flight and hotel bookings. He calls that a product problem, not a technology one. The rest is a year of reverse engineering how consumer memory systems are built, all of it his reading from the outside rather than anything documented. By his account ChatGPT went from a user managed list of facts to a running profile rebuilt in the background, roughly 4,000 tokens of dense keyword clues he could only inspect by jailbreaking his way to it. Claude started opposite, with no profile and two retrieval tools over past conversations, then added one about a quarter the size, in full sentences, refreshed daily and visible in settings. His frame for the difference is that memory is a function of compute: a profile costs to maintain and costs again in every context window it enters, so a large profile updated rarely and a small one updated daily are two answers to one budget question. Speaker info: https://x.com/shloked, https://www.linkedin.com/in/shlokkhemani/, https://shloked.com, Timestamps: 0:00 - What memory means here: consumer personalization 2:09 - ChatGPT memory v1, a list of facts 4:08 - The running profile arrives 6:04 - The trip that never happened 6:45 - A profile you cannot look at 7:23 - Claude's first version, tools instead of a profile 8:03 - Publishing that they were opposites 9:18 - Three years of convergence 11:14 - There is no single way to do memory 12:30 - Memory is a function of compute 13:47 - Continual learning is already here 15:39 - The context problem no architecture solves 17:34 - Products that each know you separately</description>
        <lastBuildDate>Thu, 13 Aug 2026 23:40:37 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://video.ut0pia.org</generator>
        <image>
            <title>Lessons from Studying Every Memory System — Shlok Khemani, Independent</title>
            <url>https://video.ut0pia.org/lazy-static/avatars/0287a09a-aae7-4840-9843-b416426e7046.webp</url>
            <link>https://video.ut0pia.org/videos/watch/9b396638-fc9e-4ef3-8129-f4f7c480ad9d</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://video.ut0pia.org/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://video.ut0pia.org/feeds/video-comments.xml?videoId=9b396638-fc9e-4ef3-8129-f4f7c480ad9d" rel="self" type="application/rss+xml"/>
    </channel>
</rss>