AI Research OS: How Second Brain Becomes Memory for Agents (Category AI4SDLC)
I saw it. speech Paul Iusztin of Decoding AI (Author of LLM Engineer's Handbook) Louis-François Bouchard of Towards AI (Author of Building AI Systems for Production) on the AI Engineer channel: "Turn 10,994 Notes Into Memory. It happened to me that the guys were solving a problem that I often face now: research on different topics is accumulating, documents are scattered across different sources and collecting context for a new question is not very difficult. I struggle with this problem in different ways, and here the authors suggested their own, which I plan to try.
But the very statement of the problem in the authors sounds like the organization of information in a form that will be convenient for reuse in research. Codex, Claude Code, ChatGPT, or NotebookLM can quickly help with a single task, but if each new session recreates the sources, findings, and open questions, the knowledge of the study does not accumulate.
To solve this problem, the authors created a projectAI Research OS" Essentially, it’s a local layer of memory between your second brain and agent-bound bandage. (agent harness). Not a regular chat with a large context window or a heavy RAG infrastructure with a vector database. The idea is more mundane: regular files, Markdown, index.yamlRaw sources and a wiki layer that an agent can read, top up and check.
The architecture is simple and boring:
- At the entrance - Obsidian, Readwise, NotebookLM, GitHub repositories, YouTube transcripts, links, PDF and local files The system adds immutable raw sources, builds an index, and generates wikis from the top: source pages, concepts, entities, comparisons, overview, synthesis, open questions and log The agent first reads the index, then the short wiki page, then the derivative pages and only if necessary climbs into the full raw document.
Interestingly, Paul notes that his second brain with notes in Obsidian works as a long-term memory and immutable snapshot. (LLM should not rewrite personal notes). Instead, a separate research-wiki is created for a specific project. In the end, we have Second brain as an archive Project wiki as a working memory, from which you can then write an article, make slides, disassemble the code base or continue research in a month.
The authors compare their approach (through simple file storage in turnip) others NotebookLM is useful for reading a set of sources, but the authors find it less convenient for agent-native and coding scripts. Vector database and full-fledged RAG are needed in product systems, but for a personal research OS it is often too much infrastructure. (By the way, I'm digging into my pet project right now, but there really are a lot of moving parts.)
If you decide to try the authors’ approach, the implementation is open on GitHub under the MIT license. The repository has skills. /research, /research-distill, /research-lint, /research-renderExamples with deep research, ingest GitHub repositories, and ingest common links. According to README, you need to start. uvClaude Code or Codex, and external CLIs for Obsidian, Readwise and NotebookLM are connected as needed. For the first experiments with repositories and links, you do not need to immediately drag the entire personal archive.
Separately, the authors say that in the project they did not plan to build SaaS and there is room for improvement: there are not enough connectors like Google Drive, Notion and Slack, less developed source provenance, assessment of source obsolescence and memory compaction. But the project shows how the system works, where you can intervene and what to adapt to your process.
P.S. I liked the authors’ presentation and I think I’ll touch their system and see how it works for my scripts.
#AI #AI4SDLC #Engineering #Research #Software #Architecture