Turn 10,994 Notes Into Memory - Paul Iusztin, Decoding AI & Louis-François Bouchard, Towards AI

Key Concepts

• AI Research OS: A personalized, file-based system designed to manage research, notes, and knowledge retrieval using LLMs.
• Second Brain: A digital repository (Obsidian, Readwise, Notion) containing an individual's accumulated knowledge.
• Deep Research Algorithm: An iterative process where an orchestrator agent generates queries, gathers resources, and synthesizes information.
• Wiki Layer: A structured, LLM-generated knowledge base that sits on top of raw data, containing comparisons, concepts, and entities.
• Context Compaction: Techniques used to summarize and condense large amounts of information to fit within LLM context windows efficiently.
• Index.yaml: A central catalog file that acts as an entry point for agents to navigate raw sources and wiki derivatives.

---

1. The Problem: Research Fragmentation

The speakers, Paul Yushin and Luis François, identify a common bottleneck in knowledge work: the "graveyard" of saved links, scattered notes, and the inability to recall or synthesize information across different platforms (Obsidian, GitHub, Google Drive). Existing tools like NotebookLM or ChatGPT are either not "agent-native," lack personalization, or fail to maintain long-term memory across different projects.

2. The Solution: AI Research OS

The proposed system is a file-based, local-first architecture that avoids the complexity of vector databases or knowledge graphs in favor of simple, inspectable Markdown files.

The Three-Layer Architecture

1. Raw Content Layer: Immutable source files (PDFs, transcripts, code, notes).
2. Index Layer: An index.yaml file that maps sources to their summaries, metadata, and wiki derivatives.
3. Wiki Layer: A synthesized, living knowledge base created by the LLM, containing:
   • Comparisons: Analysis between different concepts or architectures.
   • Concepts: Extracted definitions and summaries.
   • Entities: Key terms and their relationships.

3. Step-by-Step Methodology

The system operates through a repeatable "Deep Research" loop:

1. Input: The user provides a topic or a set of "seed" links.
2. Orchestration: A main agent breaks the topic into multiple sub-questions.
3. Retrieval: Agents query the user's "Second Brain" (Obsidian, etc.) and the public web.
4. Ranking & Scraping: A ranking algorithm filters for high-signal information, scraping only the most relevant content.
5. Synthesis: The system generates executive summaries and updates the index.yaml and Wiki files.
6. Evolution: As the user asks follow-up questions, the Wiki is updated, creating a "living" record of the research process.

4. Key Arguments & Perspectives

• Avoid Over-Engineering: The speakers argue against complex infrastructure (vector DBs, semantic search) for personal use. They advocate for a "file-based" approach because it is human-readable, easy to debug, and leverages the existing file system.
• Context Management over Context Window: The bottleneck is not the size of the context window, but how effectively the system manages memory and references. By using executive summaries and a hierarchical index, the system remains token-efficient.
• Personalization: The system is designed to be "anchored" in the user's personal values and existing notes, ensuring the AI output reflects the user's specific context rather than generic web knowledge.

5. Notable Quotes

• "The bottleneck is how can you leverage it in the future? Meaning that with an agent, the context window becomes everything... and when you stop the conversation, it loses everything." — Luis François
• "Forget the infrastructure you think you need... all that is beautiful, but adds a lot of complexity... I want a system just based on files." — Paul Yushin

6. Real-World Applications

• Content Creation: Automating the research phase for YouTube videos or articles by referencing previous work to avoid duplication.
• Software Engineering: Ingesting multiple GitHub repositories to compare architectures, permission flows, and memory systems.
• Academic/Professional Research: Compiling deep research on complex topics (e.g., Agentic AI Engineering) into a structured, queryable Wiki.

7. Synthesis and Conclusion

The AI Research OS is a framework for transforming "research into work." By moving away from static, isolated notes toward a dynamic, agent-accessible Wiki, users can build a system that compounds in value over time. The system is intentionally "rough" and modular, encouraging users to add their own connectors (e.g., Notion, Slack) and adapt the code to their specific workflows. The ultimate takeaway is that effective AI research requires a structured, persistent memory layer that allows agents to reason across a user's entire history of thought and data.