TL;DR
ChatGPT is a great query layer over your notes and a terrible storage layer. Keep the source of truth in Obsidian, Notion, or plain Markdown, then point ChatGPT at a weekly snapshot inside a Project. Use it for three jobs only — retrieval, synthesis, and finding contradictions — and demand a direct quote for every claim. Before you upload anything personal, turn off Settings → Data Controls → Improve the model for everyone. If you want citation-grounded answers instead of generative synthesis, NotebookLM (free, 50 sources per notebook as of June 2026) is the better tool.
The core problem
Six months of “chatting with my second brain” leaves most people with hundreds of conversations they cannot search across, export in bulk, or trust. ChatGPT chats are not a knowledge base: there is no global search across your account, no folder structure, no backlinks, and no guarantee any individual chat survives a model or UI change. The moment you treat conversations as storage, your knowledge becomes locked inside a product you do not control.
The fix is a clean split:
- Storage stays external. Obsidian +
.mdfiles (most portable), a Notion database, or a daily journal. This is your durable, searchable, exportable source of truth. - ChatGPT is the lens. You feed it a snapshot and use it to retrieve, synthesize, and surface connections you would have missed by hand.
If you do not have a notes habit yet, build that first. AI cannot summarize content that does not exist. A reasonable floor before this workflow pays off: roughly 50 Markdown files, a Notion workspace you actually open, or a journal more than a month old.
When this is the wrong tool
| Situation | Use this instead |
|---|---|
| No notes habit yet | Obsidian or Notion first; come back later |
| Sensitive personal data (mental health, finance, medical) | A local-only vault + offline AI plugin |
| Team knowledge sharing | A shared wiki; ChatGPT Projects are per-user, not collaborative |
| Real-time queries against constantly changing notes | Native search in your notes app; you re-upload weekly at best |
| You want citations, not generative synthesis | NotebookLM |
Before you start
- Pick one source of truth and commit. Obsidian + Markdown is the most portable. Notion works if you accept its export friction (large workspaces export slowly and split into multiple files).
- Set a snapshot cadence. Weekly is the floor, monthly is the realistic ceiling. A stale snapshot is the single biggest cause of confident, wrong “recent thinking” answers.
- Write a 200-word self-description of the recurring topics, projects, and people in your notes. This goes into Project Instructions so ChatGPT stops asking who “Sarah” is in every chat.
- Turn off training before you upload anything personal. Go to Settings → Data Controls → Improve the model for everyone and switch it off. As of June 2026 this is on by default for Free, Plus, and Pro personal accounts, and the setting applies account-wide. For one-off questions, Temporary Chat (top of the model picker) skips history, memory, and training entirely.
Step by step
-
Export a weekly snapshot of your notes as plain Markdown. Name it with an ISO date —
snapshot-2026-06-06.md— so date arithmetic and diffs are obvious later. In Obsidian, the Export to Markdown community plugin bundles a vault into one file; in Notion, use Settings → Export (Markdown & CSV). -
Create a Project called “PKM” and upload the latest snapshot. ChatGPT Projects let you attach files and persistent instructions that apply to every chat inside the Project. As of June 2026, Free accounts can store about 5 files per Project and Plus about 25; you can upload up to 10 files in a single action, each up to 512MB and 2M tokens. A single combined snapshot file sidesteps all of these caps.
-
Write the Project Instructions as a strict contract:
You are a retrieval layer over my notes, not a creative writer. - Never invent connections. If you cannot find a quote, say "Not in the notes." - Every claim must include a direct quote, with the note title and date. - Dates and titles are facts, not guesses. Do not fabricate them. - Do not include: dream journals, draft emails, anything marked sketch. -
Use the Project for exactly three jobs:
- Retrieval — “Find every note where I mentioned LangGraph, with dates.”
- Synthesis — “What have I been thinking about hiring lately? Quote each note.”
- Contradiction — “Two notes from different months that disagree on the same topic. Show both with dates.”
-
Write the output back into your notes app after any useful session, as a new note. Tag the source
#chatgpt-synth-2026-06-06so the synthetic note is auditable and never confused with original thinking. -
Run a monthly review: “What topics have I been writing about that went nowhere? Suggest three next steps for each.”
The monthly review loop
This is the workflow that actually compounds. Once a month: export → upload the new snapshot → ask for connections across the month → write three short synthesis notes back into Obsidian → ask for next-month focus areas → archive the chat with a clear name like pkm-review-2026-06.md. Diffing this month’s answers against last month’s is the real signal — it shows you what you keep circling and what you have quietly dropped.
Quality control: trust nothing without a quote
The failure mode of this entire workflow is fabricated connections. Two safeguards catch almost all of them:
- Demand both sides of every connection. For any “these two notes relate” claim, require a direct quote from each note. If the model cannot produce both, the connection is fiction — this is where hallucination concentrates.
- Verify dates and titles first. These get invented far more often than body text. Spot-check two or three against your real vault before you trust the rest.
A useful third prompt is “What’s missing?” — what topics would a human reader expect that the model did not surface? If your snapshot is well-curated, that list should be short.
NotebookLM vs ChatGPT: pick the right lens
Both can answer questions over your notes, but they are good at different things.
| ChatGPT (Projects) | NotebookLM | |
|---|---|---|
| Strength | Generative synthesis, follow-up dialog, rewriting | Citation-grounded retrieval; every answer links to the source |
| Sources | ~5 files Free / ~25 Plus per Project (June 2026) | 50 sources/notebook Free, 100 Plus, 300 Pro (June 2026) |
| Citations | You must demand quotes manually | Inline citations by default |
| Hallucination risk | Higher — synthesizes freely | Lower — grounded to uploaded sources |
| Price | Free / Plus $20 / Pro $100–$200 | Free; Plus via Google AI Plus $7.99, Pro via Google AI Pro $19.99 |
Rule of thumb: NotebookLM when you need to trust the answer, ChatGPT when you need to think with it. Many people run both — NotebookLM to find and cite, ChatGPT to synthesize and draft. NotebookLM cannot be bought standalone; it ships as a benefit of a Google AI or Workspace plan.
What about Claude? Claude Projects (bundled with Claude Pro at $20/month) handle long single documents well thanks to a 1M-token context window on Sonnet 4.6 and Opus 4.7. For the snapshot-and-query notes workflow, ChatGPT’s Project UX is still the smoother fit, but Claude is a strong second choice if you live in that ecosystem.
Privacy: the part most guides skip
If your notes contain anything you would not post publicly, do not rely on a settings toggle alone.
- Keep two vaults. A “shareable” vault you export and upload, and a private vault that never leaves your disk. Local plugins like Smart Connections or Obsidian Copilot can search the private vault without anything leaving your machine.
- Redact before upload. Run find-and-replace to strip real names, locations, and identifying details from the snapshot before it goes to OpenAI.
- Use Temporary Chat for any sensitive one-off question so it never enters memory or history.
Common mistakes
- Treating chats as storage. Conversations are not searchable across your account, not exportable in bulk, and not durably yours.
- A stale snapshot. The Project reasons over months-old data and confidently narrates “your recent thinking.” Update on schedule.
- Skipping the training toggle before uploading personal notes to a default account.
- Accepting unquoted connections. Always require a direct quote, note title, and date.
- Pinning more than five chats. The pin list becomes useless the moment it grows past one screen.
- Mixing work and personal notes in the same Project. The cross-talk produces strange recommendations; use separate Projects.
FAQ
- Isn’t this just RAG? In spirit, yes, but done by hand. A real retrieval-augmented setup (NotebookLM, or a custom GPT with file search) automates the retrieval and grounds answers to citations. The DIY snapshot approach trades that automation for control over exactly what the model sees.
- NotebookLM or this? NotebookLM wins on citation-grounded retrieval and source limits (50 per notebook free as of June 2026). ChatGPT wins on synthesis, generative drafting, and follow-up dialog. They are complementary — use NotebookLM to find and cite, ChatGPT to think and write.
- How many files can a ChatGPT Project hold? As of June 2026, roughly 5 on Free and 25 on Plus, with up to 10 uploaded per action and a 512MB / 2M-token cap per file. Combining your vault into a single snapshot file avoids the per-file count entirely.
- Will OpenAI train on my notes? Only if you leave training on. Turn off Settings → Data Controls → Improve the model for everyone (on by default for personal accounts), or use Temporary Chat. Opting out does not delete existing history; it stops new chats from being used for training.
- Why not just paste notes into the chat? A Project persists files and instructions across every chat inside it. Pasted text lives in one conversation, never gets searched again, and disappears from your workflow the moment that chat scrolls away.