{ "video": { "id": "ZRM_TfEZcIo", "title": "Turn 10,994 Notes Into Memory - Paul Iusztin, Decoding AI & Louis-François Bouchard, Towards AI", "duration": 2372, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "Online" }, "analysis": { "video_id": "ZRM_TfEZcIo", "title": "Turn 10,994 Notes Into Memory - Paul Iusztin, Decoding AI & Louis-François Bouchard, Towards AI", "one_liner": "Paul Iusztin and Louis-François Bouchard show how they wired ~10,000 Obsidian/Readwise notes into a file-based \"AI research OS\" — a deep-research loop aimed at your own second brain that writes a wiki of raw sources, an index.yaml and LLM-generated derivatives, queried by Claude Code/Codex skills with no vector database anywhere.", "summary": "Paul has 5,000+ notes in Obsidian, another 5,000 in Readwise and more scattered in Notion and Google Drive, growing 250 files a month, and can never recall the right ones when he starts an article, video or codebase. The talk argues the bottleneck is not how much context you hand a model but whether you can leverage it again next session — so they built a system of plain markdown files that sits between your second brain and harnesses like Claude Code and Codex. They walk through three versions: v1 deep research over hand-picked golden links producing a flat research.md, v2 the same loop aimed at the second brain instead of just the public web, and v3 adding a persistent wiki layer (raw/ + wiki/ + index.yaml) that keeps growing as you ask it questions. Three live demos run the open-source AI Research OS plugin over an article brain-dump, three harness repos, and three random links.", "key_points": [ "Explicit tool decision tree: quick question → Google/ChatGPT/Claude; one-off small change or single article → Codex or Claude Code; work you'll keep digging into → a system where sources stick around. NotebookLM is rejected because you don't own it, can't personalize it, it isn't agent-native and it's weak for coding; RAG with vector databases is rejected for personal use because it needs infrastructure and is hard to inspect or edit by hand.", "The stated bottleneck: \"with an agent, the context window becomes everything, the database, the file system, the memory, the reasoning space\" — and it's all lost when the conversation ends. More context isn't the fix; proper memory and context management is.", "v1 deep research algorithm: scrape hand-picked golden links first as seed context for framing questions, then an orchestrator agent spawns sub-agents that each own one question and use Gemini grounded in Google; each returns links plus executive summaries, and the main agent aggregates so context doesn't explode. Three rounds × six queries ≈ 40–50 links, then a ranking pass scores each source against the topic, fully scrapes only the top-K and keeps summaries for the rest. It produced 35 course lessons quickly.", "v2 aims the same loop at the second brain (Obsidian, Readwise, NotebookLM, GitHub) plus the public web, so the input is just a topic — the golden links are already a reflection of your second brain, no manual curation needed.", "v3 fixes the fact that a research.md is \"a pile of static data\" you'd have to regenerate from scratch (expensive in tokens and time) by adding a wiki layer: immutable raw/ files, an LLM-generated wiki/ of sources, concepts, entities, comparisons, notes and open questions, and an index.yaml catalog with per-source summaries and metadata (origin, title, authors, publication date).", "\"Forget the infrastructure you think you need\" — no vector DB, no knowledge graph, no semantic or text search. Retrieval is a token-efficient reference hierarchy: index.yaml → the source's expanded executive summary → wiki derivatives → and only as a last resort the full raw source.", "The wiki is alive: every question can create a new concept, note or comparison file and is written to a log, so it evolves from conversation, not just ingestion — \"a true reflection of yourself, of what you haven't understood.\"", "The wiki does not sit on the whole second brain. Paul uses Tiago Forte's PARA method with everything piped into a flat resources list; Obsidian stays an immutable snapshot the LLM never writes to, and each project scopes down its own wiki from it.", "Demos: (1) the research skill pointed at a brain-dump file with depth modes light (1 round × 3 queries), fast (2 rounds × 3 queries) and deep, taking 10–20 minutes and auto-generating comparisons like \"agentic RAG vs file systems\" and \"compaction vs recursive language models\"; (2) ingesting the OpenCode, Pi and Hermes repos with no deep research at all to compare architectures, memory systems and permission flows — useful if you want to write your own harness; (3) three random links. Demos 2 and 3 need only git and curl, no Obsidian or Readwise setup.", "Louis-François consolidated everything into Obsidian — Granola meeting recaps auto-saved, Apple Notes, tweets from saved Chrome tabs, his Claude Code skills folder — using Codex to set up the automation; he added YouTube transcript ingestion with one prompt in a few seconds." ], "takeaways": [ "Stop pasting the same links and PDFs into a fresh Codex/Claude Code session: put your sources in plain markdown on the file system, with an index the agent reads as its entry point, so the research compounds across sessions and projects.", "Build the retrieval hierarchy instead of the infrastructure — index summary, then executive summary, then derivative, then raw file — and only pay for the full source when the cheaper layers fail.", "Point your deep research loop at your own notes rather than only the public web; your saved articles and highlights already are the golden links, so the input collapses to a topic.", "Keep the LLM out of your hand-written notes: treat the vault as an immutable snapshot and let each project derive its own scoped wiki from it.", "Default to the light or fast depth setting — the deep research loop burns a lot of tokens and 10–20 minutes per run; save 'deep' for when you genuinely need to sweep tons of notes.", "Clone the AI Research OS workshop repo as Claude Code/Codex plugins and add the connectors you actually need (Google Drive, Notion, Slack are missing on purpose) — the authors say a connector like YouTube transcripts is a one-prompt job." ], "topics": [ "memory", "context-management", "deep-research", "agents", "second-brain", "rag", "knowledge-base", "agent-harnesses" ], "tools": [ "Obsidian", "Readwise", "NotebookLM", "Notion", "Google Drive", "Claude Code", "Codex", "ChatGPT", "Gemini (grounded in Google Search)", "Gemini Deep Research", "Granola", "Apple Notes", "Bright Data", "GitHub", "YouTube", "OpenCode", "Pi", "Hermes", "MCP", "Slack", "AI Research OS (workshop repo)", "Decoding AI", "Towards AI / Towards AI Academy", "What's AI", "LLM Engineer's Handbook", "Building AI Systems for Production", "PARA method (Tiago Forte)" ], "quotes": [ { "text": "I'm always losing my research. For example, my reading list is a graveyard.", "at": "00:52", "url": "https://www.youtube.com/watch?v=ZRM_TfEZcIo&t=52s" }, { "text": "The bottleneck is how can you leverage it in the future? Meaning that with an agent, the context window becomes everything, the database, the file system, the memory, the reasoning space. It has to do it all, and when you stop the conversation, it loses everything.", "at": "08:46", "url": "https://www.youtube.com/watch?v=ZRM_TfEZcIo&t=526s" }, { "text": "What I want to start with is that you should forget the infrastructure you think you need, such as vector databases, knowledge graphs, semantic search, text search.", "at": "19:38", "url": "https://www.youtube.com/watch?v=ZRM_TfEZcIo&t=1178s" }, { "text": "The project is the work, and your second brain is the research.", "at": "27:00", "url": "https://www.youtube.com/watch?v=ZRM_TfEZcIo&t=1620s" } ], "words": 7562 }, "summary_url": "/#ZRM_TfEZcIo", "transcript": { "html": "/transcripts/ZRM_TfEZcIo.html", "txt": "/transcripts/ZRM_TfEZcIo.txt", "vtt": "/transcripts/ZRM_TfEZcIo.vtt" } }