{
  "video": {
    "id": "9_oy4owGtPI",
    "title": "Agent Memory Is Solved. Agent Learning Isn't. — Karthik Ranganathan, Yugabyte",
    "duration": 1236,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "9_oy4owGtPI",
    "title": "Agent Memory Is Solved. Agent Learning Isn't. — Karthik Ranganathan, Yugabyte",
    "one_liner": "Per-agent memory is solved, but learning across agents isn't: agents hand off outputs (an MD file, a PDF) without the reasoning or dead ends behind them, so Yugabyte built Mako, a shared persistence layer where agents push context into a private space and promote it, under human supervision and with traceable provenance, into a shared one.",
    "summary": "Karthik Ranganathan, co-founder and CEO of Yugabyte, says the company spent almost a year failing to ship four agents of its own. The cause was a data problem: when one agent hands work to another, the second agent gets only the output, not the reasoning, dead ends or context, so it re-derives everything and burns tokens doing it. He argues that more memory, shared state, RAG over transcripts, fine-tuning and bigger context windows are all architecture gaps, not learning. What works is shared context that only becomes durable through supervised promotion and can be traced back to its source. Heather's demo of Mako's 'data packs' showed a session resumed across chats, private memory promoted so a different agent (Codex) could find it, and embedded traces used to check token spend.",
    "key_points": [
      "Yugabyte runs four agents: two internal (Hagen for customer support, Growth Vector for coordinated sales and marketing go-to-market) and two for customers (Perf Advisor, an agentic database admin, and Voyager, for migrating to a cloud-native database). They tried for almost a year and failed to ship any of them before realizing the failure was a data problem.",
      "The core failure is lost context at handoff. Agent B gets only agent A's output (an MD file, PDF or document), not the reasoning, dead ends or failures behind it, so it re-derives all of that at its own token cost. This is like a new team member re-proposing ideas the team already rejected weeks ago.",
      "Five misconceptions, each called an architecture gap: agent memory is not learning, just more (and more expensive) memory; shared state is not shared knowledge; RAG over everything is not learning or quality; fine-tuning does not scale for persisting lessons; and a bigger context window is the quickest, most expensive route to failure because context rots.",
      "Three requirements: a context agent A can share with agent B; supervised promotion before anything becomes durable, because context can degrade and not everyone should contribute ('just like code'); and provenance, so every fact traces back to a data source, person or conversation.",
      "Approaches that failed: a scalable distributed RAG knowledge base; a tiny Postgres per agent, where every repeatable workflow got reinvented on the fly; and a backend with no UI, where they couldn't tell what was going on.",
      "Their first RAG pipeline, built from the best academic components, scored 14% faithfulness on a public MD benchmark, 20% on a PDF one and 57% context precision. Tracing and tuning one variable at a time raised these to 65, 74 and 82, while context fell from over 7,000 chunks to about 1,000 (a seventh of the size), cutting token spend.",
      "Mako is an 'agent-native persistence layer' exposed as a data pack. It combines vector, graph, relational and NoSQL storage and holds knowledge base, memory, conversations and observability. Agents push context into a private space and promote it to a shared space that other agents and humans retrieve from. The deck for this talk was built that way: Heather sent a data pack share request, Karthik built the slides in Claude Desktop without a meeting, and one image came from Claude Design.",
      "Demo (through an MCP server and a skill): Claude stored an incident fact (signup emails failed because they were on the SES sandbox, not prod) with data pack create and memory add, then recalled it from a fresh chat outside Projects with memory turned off. Codex could see the data pack but not the fact until that one memory was promoted, at fine granularity, to the shared space. Embedded traces show the decision trace down to each tool call, so token savings can be checked."
    ],
    "takeaways": [
      "Hand off reasoning and dead ends, not just the output artifact, between agents (and between people). Otherwise the receiving agent pays in tokens to re-derive them.",
      "Gate shared agent knowledge with human-supervised promotion, as you would with code review, and keep provenance for every fact so a bad piece of context can be traced to its source and fixed.",
      "Before scaling a RAG pipeline, trace it and tune one variable at a time. Measure faithfulness and context precision; fewer, better chunks can raise quality and cut cost together.",
      "Keep a human-visible UI and human-in-the-loop over agent data infrastructure. Human promotion decisions double as labelled examples for training self-improving orchestration later.",
      "Share state about incidents (for example, 'this server is down') in a common store so the first agent to hit an outage tells the rest, instead of every agent re-calling the same failing service."
    ],
    "topics": [
      "agents",
      "agent memory",
      "multi-agent",
      "context engineering",
      "rag",
      "mcp",
      "knowledge management",
      "token cost"
    ],
    "tools": [
      "Yugabyte DB",
      "Mako",
      "Postgres",
      "Claude Desktop",
      "Claude Design",
      "Claude Code",
      "Codex",
      "ChatGPT",
      "Claude.ai",
      "MCP",
      "Amazon SES",
      "Slack",
      "LinkedIn",
      "Discord",
      "Hagen",
      "Growth Vector",
      "Perf Advisor",
      "Voyager"
    ],
    "quotes": [
      {
        "text": "if a database company is failing at a data problem be it for agentic it becomes personal",
        "at": "02:16",
        "url": "https://www.youtube.com/watch?v=9_oy4owGtPI&t=136s"
      },
      {
        "text": "No human in the loop equals really bad results.",
        "at": "06:30",
        "url": "https://www.youtube.com/watch?v=9_oy4owGtPI&t=390s"
      },
      {
        "text": "we barely talked, right? We didn't need to. We just shared context and our agents did the work",
        "at": "09:57",
        "url": "https://www.youtube.com/watch?v=9_oy4owGtPI&t=597s"
      },
      {
        "text": "human promotions today kind of equal the training signal for self-improving orchestration tomorrow",
        "at": "16:48",
        "url": "https://www.youtube.com/watch?v=9_oy4owGtPI&t=1008s"
      }
    ],
    "words": 4353
  },
  "summary_url": "/#9_oy4owGtPI",
  "transcript": {
    "html": "/transcripts/9_oy4owGtPI.html",
    "txt": "/transcripts/9_oy4owGtPI.txt",
    "vtt": "/transcripts/9_oy4owGtPI.vtt"
  }
}