{
  "video": {
    "id": "KGFyOtl5ktI",
    "title": "Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain",
    "duration": 999,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "KGFyOtl5ktI",
    "title": "Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain",
    "one_liner": "Memory is the durable context an agent references to guide future runs. A trace only becomes memory once a lesson from it is filtered out and written back to long-term context, giving a read → run → filter → write loop that makes the next run better.",
    "summary": "Jake Broekhuizen leads the labs team at LangChain. He argues that teams have become good at observing agents but not at turning what they observe into lessons for future runs: if an agent's context never changes, today's mistake will still be there tomorrow. He groups memory into semantic, episodic and procedural types, separates working memory from long-term memory, and describes memory as a read-write cycle in which deciding what to keep is the hard part. He grounds this in a financial-services budgeting agent whose tone slipped, fixed with LangSmith observability, a new background analysis step and LangSmith's Context Hub.",
    "key_points": [
      "Running example: a financial-services agent for spending, budgeting and non-investment advice had its tone slip from advisory to directive (\"it looks like you need to cancel some subscriptions and move some money into savings\"). Fixing that by hand meant reading traces, editing context and checking for regressions, which is hard to scale.",
      "Definition: memory is \"the durable context that an agent references that will guide future runs.\" A trace, transcript or log is only evidence. It becomes memory when its lesson is turned into durable context.",
      "Taxonomy borrowed from cognitive science: semantic memory is what the agent knows (facts, preferences). Episodic memory is what it has experienced (patterns, past interactions, examples). Procedural memory is how it should behave (instructions, skills, rules).",
      "Most of the visible gains come from procedural memory. The tone fix was a new rule to follow, not a new fact or a banned word.",
      "Working (short-term) memory covers scratch pads, tool results and retrieved files for the current run. Long-term memory sits in a store. Harnesses inject it in different ways: into the prompt, through tools, from a store, through files, or by changing runtime state.",
      "The core loop is a read-write cycle, related to sleep-time compute and 'dreaming'. The agent reads long-term context into the run (e.g. progressive disclosure of skills), the run produces a trace, a filter step separates signal from history, and the signal is written back. In simpler terms: capture, analyze, update.",
      "LangSmith implementation: observability does the capture; a new background analysis step extracts signal from traces based on what matters to you; LangSmith's Context Hub stores instructions, skills, markdown files, policies and rules that the agent pulls on its next run. In the demo, the analysis found patterns and changed the skill markdown files in Context Hub, and the agent's tone changed on later runs.",
      "Gotcha: in long-running background agents, a memory update made early may never reach later runs because of caching and runtime state. Broekhuizen says he was 'banging my head on a wall' over this."
    ],
    "takeaways": [
      "Treat traces as a source of signal, not just logs. Build a filter step so that most trace data stays as history (some becomes offline eval datasets) and only a small subset changes the agent's memory.",
      "When an agent's behaviour drifts, as with tone, fix it in procedural memory (rules and instructions) rather than adding facts or banned words.",
      "Work out what can and can't be cached so that memory updates actually reach future runs, especially for long-running background agents.",
      "Choose what the agent may update itself and what needs human review. Procedural memory such as instructions, policies and tone guidelines drives about 75% of behaviour, so show proposed changes to a human before they enter the hot-reload path."
    ],
    "topics": [
      "agents",
      "agent memory",
      "observability",
      "continual learning",
      "procedural memory",
      "human in the loop",
      "context engineering"
    ],
    "tools": [
      "LangChain",
      "LangSmith",
      "LangSmith Context Hub",
      "Hermes Agent"
    ],
    "quotes": [
      {
        "text": "Agents that are observable tell you what happened. Agents that are adaptive are able to change",
        "at": "04:11",
        "url": "https://www.youtube.com/watch?v=KGFyOtl5ktI&t=251s"
      },
      {
        "text": "Not everything should become a memory update.",
        "at": "13:16",
        "url": "https://www.youtube.com/watch?v=KGFyOtl5ktI&t=796s"
      },
      {
        "text": "those are probably going to drive about 75% of the behavior of my agent.",
        "at": "15:01",
        "url": "https://www.youtube.com/watch?v=KGFyOtl5ktI&t=901s"
      },
      {
        "text": "It's how experience becomes context.",
        "at": "15:51",
        "url": "https://www.youtube.com/watch?v=KGFyOtl5ktI&t=951s"
      }
    ],
    "words": 3484
  },
  "summary_url": "/#KGFyOtl5ktI",
  "transcript": {
    "html": "/transcripts/KGFyOtl5ktI.html",
    "txt": "/transcripts/KGFyOtl5ktI.txt",
    "vtt": "/transcripts/KGFyOtl5ktI.vtt"
  }
}