{
  "video": {
    "id": "5ZGyKWjQDr0",
    "title": "Lessons from Studying Every Memory System — Shlok Khemani, Independent",
    "duration": 1171,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "5ZGyKWjQDr0",
    "title": "Lessons from Studying Every Memory System — Shlok Khemani, Independent",
    "one_liner": "After a year reverse-engineering how ChatGPT, Claude, Gemini and Poke do memory, Shlok Khemani shows that both big chat products independently abandoned RAG and converged on a 'running profile' updated in the background — and argues memory is a compute trade-off you must build in-house, not a component you outsource.",
    "summary": "Part one traces three years of consumer memory: ChatGPT's Feb 2024 v1 fact list, its April 2025 v2 'user knowledge memories' running profile, Claude's Aug 2025 tools-only design with no profile at all, and Claude's v2 profile released the same day the speaker's 'Claude's memory architecture is the opposite of ChatGPT's' post hit the Hacker News front page. Part two draws the lessons: there is no single way to do memory, so serious teams build it alongside the product rather than outsourcing it; memory is a function of compute, and ChatGPT and Claude sit on opposite ends of the maintenance-cost vs serving-cost trade-off; and running profiles already constitute continual learning, just outside the weights. His rant is that the real ceiling is context, not architecture — products aren't curious about gaps, don't reason over email or calendar, and each rebuilds its own siloed memory of you.",
    "key_points": [
      "ChatGPT memory v1 (Feb 2024) extracted user-stated facts into a visible, deletable list injected into every conversation; its two flaws were putting the burden of memory management on the user and staleness — a memory saying the speaker was going to Bengaluru is still in his context window today.",
      "ChatGPT memory v2 (April 2025) added 'user knowledge memories' — a running profile rebuilt every few days from recent conversations (the 'dreaming' process Lance Martin also spoke on), written as extremely dense keyword-like clues across 16 sections totalling almost 4,000 tokens, and not viewable in settings (the speaker read his own only by jailbreaking).",
      "Staleness survived v2: because his source conversations were about choosing between Thailand and Turkey, the profile records overlapping 2025 trips to both, and he has never been to Turkey.",
      "Claude's memory v1 (Aug 2025) was the opposite design: no profile, no fact list, just two tools — search past conversations by keyword/topic, and by time period — with every conversation starting fresh and retrieval happening on demand.",
      "Claude v2 shipped on 11 September, the same day the speaker's 'Claude's memory architecture is the opposite of ChatGPT's' post hit the Hacker News front page: a user-visible profile of ~1,000 tokens in complete sentences rather than dense keywords, updated every 24 hours, with explicit user edits that trigger a resynthesis; unchanged since.",
      "ChatGPT shipped two 2026 updates — a tool to search past conversations, and (start of June) partial profile visibility as an LLM-generated summary of a profile that is already an LLM-generated summary — plus deprecation of the v1 fact list; the two products have now converged on profile + conversation-search tools.",
      "Nobody landed on the assumed RAG answer (chunk, embed, vector store, semantic search); Gemini's profile carries per-memory timing logs, and agents like Claude Code, OpenClaw and Hermes use markdown files, heartbeats, knowledge bases and skills. Every top consumer product has memory and all of them build it in-house.",
      "Memory costs split into maintenance (update frequency × compute per update) and serving (profile length, paid on every conversation). ChatGPT takes high serving / low update cost (4,000 tokens, every few days); Claude takes the exact opposite (1,000 tokens, every 24 hours). Unconstrained you'd update hourly with Fable plus Opus sub-agents and store 400,000 tokens.",
      "Running profiles already are continual learning — profile shapes conversations, conversations feed synthesis, synthesis rewrites the profile — but the loop runs outside the weights; whether it moves into weights per-person is open, since continual learning amortizes at enterprise scale but not individual scale.",
      "The binding constraint is context, not architecture: ChatGPT could not resolve the Thailand/Turkey conflict because the decision happened in person and the confirming evidence sat in his email as flight and hotel bookings it never reasons over — and worse, it doesn't notice the conflict or try to fill the gap. That's a product problem, not an LLM limitation."
    ],
    "takeaways": [
      "Don't outsource memory to a third-party layer — build it in-house alongside the product, because the architecture has to evolve with what you're building and no single design fits.",
      "Stop reaching for RAG-over-conversations by default; a background-updated running profile injected into every context is what the leading consumer products actually converged on.",
      "Decide your memory design as an explicit compute budget: pick where you sit between ChatGPT's long-profile/infrequent-update and Claude's short-profile/daily-update trade-off, and know which cost you're paying.",
      "Design for staleness and conflict — deliberation-shaped conversations produce false facts (Thailand vs Turkey), so build detection and resolution rather than assuming extraction is enough.",
      "Make the system curious: have it notice gaps and contradictions and pull from the rich context sources users already have (email, calendar, photos) instead of only mining chat transcripts."
    ],
    "topics": [
      "memory",
      "personalization",
      "consumer-ai",
      "context-engineering",
      "continual-learning",
      "rag",
      "agents"
    ],
    "tools": [
      "ChatGPT",
      "GPT-4",
      "Claude",
      "Claude Code",
      "Gemini",
      "Poke",
      "OpenClaw",
      "Hermes",
      "Fable",
      "Opus",
      "Hacker News",
      "Gwern's 'Guardian Angels' essay"
    ],
    "quotes": [
      {
        "text": "The point being that there is no one way to do memory. The implications of this is that memory cannot be outsourced. If you're a serious team, you do not outsource memory.",
        "at": "11:31",
        "url": "https://www.youtube.com/watch?v=5ZGyKWjQDr0&t=691s"
      },
      {
        "text": "this is what I mean by memory is a function of compute. You have to really think about how much compute you want to put into memory.",
        "at": "13:30",
        "url": "https://www.youtube.com/watch?v=5ZGyKWjQDr0&t=810s"
      },
      {
        "text": "Yet your memory system is capped by how much context it can gather about you.",
        "at": "15:55",
        "url": "https://www.youtube.com/watch?v=5ZGyKWjQDr0&t=955s"
      },
      {
        "text": "So for me, none of this feels like 2026. And what I keep asking myself every day is when will personal AI feel like personal AI?",
        "at": "18:21",
        "url": "https://www.youtube.com/watch?v=5ZGyKWjQDr0&t=1101s"
      }
    ],
    "words": 3456
  },
  "summary_url": "/#5ZGyKWjQDr0",
  "transcript": {
    "html": "/transcripts/5ZGyKWjQDr0.html",
    "txt": "/transcripts/5ZGyKWjQDr0.txt",
    "vtt": "/transcripts/5ZGyKWjQDr0.vtt"
  }
}