A personal research system can turn a large Obsidian and Readwise archive into context that agents can query and update.
2
A file-based architecture with raw sources, an index, and a wiki reduces the amount of content an agent must read while keeping the data inspectable.
3
Deep research becomes more useful when it searches both the public web and a person's curated second brain, then stores the results as reusable project memory.
Summary
Paul Iusztin and Louis-François Bouchard show how they turned years of notes, videos, documents, and code repositories into a personal AI research OS. Their system sits between an agent harness such as Codex or Claude Code and a local second brain stored mainly in Obsidian. It runs deep research over personal sources and the public web, saves raw files, builds an index, and creates a wiki of summaries, concepts, entities, and comparisons. The wiki lets an agent start with lightweight references and read a full source only when needed. The authors explain why NotebookLM is hard to personalize and why vector databases add too much infrastructure for their daily workflow. They also demonstrate research on agent engineering, ingestion of GitHub repositories, and custom web links. The repository is deliberately rough and builder-oriented. Future work includes more connectors, memory compaction, source provenance, and better source ranking.
A large second brain is difficult to turn into useful working context
Paul Iusztin has more than 5,000 notes in Obsidian, another 5,000 in Readwise, and additional material in Notion and Google Drive. The archive grows by about 250 files per month. When he starts an article, project, codebase, or feature, he wants to retrieve only the notes with useful signal. His reading list has become a graveyard of saved posts, articles, videos, and repositories. The problem is not collecting information. It is remembering what exists and finding material that applies to the work in front of him.
The right research tool depends on how persistent and complex the work is
Louis-François Bouchard separates quick questions from work that needs a long-lived research trail. For a simple answer, he recommends Google, ChatGPT, Claude, or another fast tool. Codex or Claude Code fits a small repository change or a one-off article. When research must remain available for later questions, future videos, or repeated projects, a more persistent system is needed. NotebookLM can digest sources well, but Bouchard says it is difficult to personalize, is not agent-native, and is weak for coding tasks because it is browser-based.
The system keeps personal knowledge outside the agent's direct control
Bouchard moved meeting recaps, personal notes, saved articles, and other material into Obsidian, where the files remain local and usable across his phone and computers. The AI Research OS repository connects this local material with Readwise, NotebookLM, GitHub, YouTube links, web links, and documents. The authors present the repository as a set of skills and plugins for Codex and Claude Code that users can adapt. It is designed to work with an existing archive rather than requiring someone to begin with an empty knowledge base.
Deep research searches personal sources alongside the open web
The first version accepted a topic and manually selected golden links, then used an orchestrator and multiple agents to ask questions, search Google with Gemini grounding, summarize links, rank sources, and fully scrape only the highest-ranked results. After three rounds with six queries per round, the system could produce roughly 40 to 50 links, which created noise. The second version added the authors' own sources, including Obsidian, Readwise, NotebookLM, and GitHub. The topic itself became enough to seed searches, because the curated second brain supplied many of the useful sources.
A wiki layer makes research editable instead of disposable
The static research Markdown file worked for generating course lessons, but it became expensive to rebuild whenever a question changed or information became stale. Version three stores each source as a raw file, creates an index, and generates a wiki on top. The wiki can contain source summaries, comparisons, concepts, entities, repository notes, and open questions. New links and new research rounds can be added later. Questions also leave traces by creating new notes or comparisons, so the knowledge base changes through use rather than only through ingestion.
Plain files and references can replace a database for this personal workflow
Iusztin argues that a personal research OS does not need the infrastructure of a production retrieval system. The raw folder contains immutable source data. An index.yaml file catalogs sources, summaries, metadata, and links to derived wiki pages. The wiki contains LLM-created derivatives. An agent begins with the index, follows a source-level summary, then follows links to concepts, entities, notes, or comparisons. It reads the complete article, paper, video, or other raw source only when the lighter layers do not answer the question.
The wiki is scoped to projects while the original second brain stays unchanged
Iusztin uses the PARA method to organize his broader Obsidian archive into projects, areas, resources, and archive. The second brain remains an immutable snapshot that the language model does not edit. When he starts a project, such as an article, video, slide deck, book, course, or codebase, the research loop selects and transforms relevant material into project-specific memory. This keeps personal notes under his control while still allowing agents to use them as research.
The demos show research over notes, repositories, and ordinary links
The first demo researches agentic harness engineering from a topic file and selected references, then offers light, fast, and deep research modes that control query and round counts. The resulting Obsidian wiki contains raw files, an index, source summaries, comparisons, concepts, and entities. The second demo clones repositories for OpenCode, Pi, and Hermes, then creates notes about architecture, subagents, memory, and permissions and compares them. The third demo ingests custom URLs without requiring Obsidian or Readwise, using Git and curl.
The project remains a teaching repository with known limits
Bouchard says the system still needs connectors for services such as Google Drive, Notion, and Slack. It also has weaknesses in identifying outdated, weak, or strong sources. The workflow runs through Codex or Claude Code in a terminal, and the lack of a polished interface is intentional because the goal is to teach memory and context management. Planned improvements include stronger linting, better memory compaction, source provenance, and improved source ranking. The authors also describe a longer Agent Engineering course built around a related multi-agent research and writing system.
"You need a proper memory and context management, and ideally some personality with it."08:59
Who should watch
You have a large collection of notes and saved links, but repeatedly search for the same material when starting a new project.
You want agents to use personal research without putting your entire archive into a vector database or letting a model edit the original files.
You are building research or writing agents and need examples of source ingestion, indexing, summaries, comparisons, and project-scoped memory.
