How to Build an AI Study Assistant That Knows Your Notes
For students and researchers · Based on Matt Giaro AI Second Brain Build
// TL;DR
Students and researchers drown in lecture notes, papers, and textbook highlights that never connect. This framework builds an AI study assistant in Obsidian and Codex that ingests all your material into a cross-linked wiki, then answers questions using only what you've actually studied. Replace the CRM pillar with a Classroom Notes pillar, clip research papers via the Web Clipper, and paste lecture notes into RAW. Ask 'What did I save about neuroplasticity?' and get a grounded response citing your own sources — never a generic answer. Use it to study smarter and write reflection journals grounded in your real research.
Why do lecture notes and papers never connect?
You take diligent lecture notes, highlight key papers, and annotate textbooks — but each lives in its own silo. When exam season hits, you can't see how the concept from week three relates to the paper you read in week nine. This is the Dumping Ground Problem applied to studying: everything's stored, nothing resurfaces.
The AI Second Brain Build solves it by processing all your study material into a single cross-linked wiki, extracting concepts, authors, and themes, and letting you query across everything you've learned at once.
How do you adapt the three pillars for studying?
Keep the Wiki/Knowledge Base as the gravitational core, but replace the CRM pillar with a Classroom Notes pillar. The three pillars become: Wiki (your central knowledge store), Classroom Notes (structured per-course or per-lecture records), and Journal (study reflection logs).
Setup is straightforward:
1. Install Obsidian, create a vault, and note the folder path.
2. Add the Obsidian Web Clipper to clip research papers and articles into RAW — it even captures full YouTube lecture transcripts.
3. Paste or type your own lecture notes directly into RAW as markdown files.
4. Open the vault in Codex, build the Karpathy wiki architecture, and configure agents.md to extract concepts, authors, and themes.
How do you ask questions answered only from your own material?
This is the whole point. When you ask 'What did I save about neuroplasticity?' in Codex, the system gives a Grounded Response that cites only your ingested sources — the paper you read, the lecture you attended, the video you clipped. It won't pad the answer with generic Wikipedia-level knowledge the LLM invents.
That matters for academic integrity and for actual learning: you're reinforcing what you've genuinely engaged with, and you can trace every claim back to a source in your wiki. The RAW → Processed pipeline also tells you exactly which readings you've ingested and which you still haven't touched.
How do study reflection journals work?
Start a chat with the trigger word `journal` and reflect on a study session: 'journal: I struggled to understand backpropagation again today.' The journal grounds its response in your saved research on the topic and logs the entry.
Over weeks, Pattern Detection surfaces recurring struggles — if backpropagation keeps appearing, the system flags it and suggests you revisit the specific sources you saved on it. Your study journal becomes a map of your weak spots backed by your own material.
How do you keep everything organised and backed up?
Set up the hourly automation in Codex so new papers and notes get processed without manual prompting — clip a paper, and by the next hour it's wiki-fied and cross-linked. Connect a private GitHub repo so your entire semester's knowledge is backed up every hour.
Use the Obsidian graph view to see how concepts across courses interlink. A dense graph means you're building genuine interdisciplinary understanding, not isolated facts.
Next step: Create an Obsidian vault for your current course, clip three key papers and paste one set of lecture notes into RAW, then run your first processing pass in Codex and ask it to summarise what you've studied.
// FREQUENTLY ASKED QUESTIONS
Can I paste my own lecture notes, or only clip web content?
Both. The Obsidian Web Clipper handles research papers, articles, and YouTube lectures with full transcripts, but you can also paste or type your own lecture notes directly into the RAW folder as markdown files. The AI processes all sources in RAW the same way — extracting concepts, authors, and themes and cross-linking them into the wiki regardless of origin.
Will it make up answers from generic AI knowledge?
No, if configured correctly. The agents.md rules require Grounded Responses anchored to your saved content, not blank-slate LLM knowledge. When you ask what you've saved about a topic, it cites only your ingested sources. This is ideal for study because every answer traces back to material you've actually engaged with, supporting both learning and academic integrity.
How does this help with exam revision specifically?
It lets you query across an entire term's material at once and see cross-links between concepts you studied weeks apart. The graph view reveals connections you'd otherwise miss. Your study journal's Pattern Detection also surfaces recurring weak spots, pointing you to the exact sources you saved on those topics so revision targets your real gaps.