TL;DR
Notes you ask ChatGPT to “remember” disappear into chat history within a day. The fix is a single Markdown file (research-log.md) that lives inside a ChatGPT Project, gets re-uploaded after every reading session, and feeds back into the model next time. The Project gives you persistent memory across chats; the file gives you a stable source of truth ChatGPT can re-read. Below is the exact loop, the current per-plan file caps (as of June 2026), and the mistakes that quietly break compounding.
Why chat history is the wrong place for notes
A normal ChatGPT thread is a dead end for research. Two reasons:
- The context window is finite. GPT-5.5 Instant carries roughly 32K tokens of context on Plus and Business; GPT-5.5 Thinking goes up to ~256K tokens when you select it. That sounds large, but a long literature-review thread blows past it, and once it does, the model silently forgets your earliest sources mid-conversation.
- Threads do not talk to each other. Notes you took in Monday’s chat are invisible to Thursday’s chat. There is no synthesis across sources unless every source lives in one re-readable place.
A ChatGPT Project solves the second problem (it keeps memory and files scoped to one research topic), and a single uploaded file solves the first (the model re-reads it instead of relying on a window that keeps scrolling away).
What you need first
- A ChatGPT account. Projects are available on Free, Go ($8), Plus ($20), and Pro ($100/$200) as of June 2026 — Projects became free in early 2026.
- One plain-text or Markdown file you will treat as the master log:
research-log.md. - A topic with enough sources (3+ papers or reports a week) that compounding actually pays off. For a single paper summary, skip all of this and just paste the PDF into one chat.
The loop, step by step
- Create one Project per research topic. Name it after the question, not the field (“Does spaced repetition help adult L2 vocab?” beats “Language Learning”). Add
research-log.mdto the Project as its anchor file. - Open a fresh chat inside the Project for each new source. Paste the source (PDF, URL, or excerpt) and one line of intent: “I’m assessing whether this supports or undercuts the spaced-repetition claim.”
- Ask for a structured extract, not a summary. Have ChatGPT return: (1) a one-paragraph plain-language summary, (2) three key claims as direct quotes with page numbers, (3) one or two open questions the source raises, (4) one paragraph on how it connects to or contradicts what is already in your log.
- Paste that block into
research-log.mdwith a header line: date, source citation, and a short tag (#spaced-rep,#contradicts-rivera). - Re-upload
research-log.mdto the Project, replacing the old copy. This is the step people forget — the Project caches the version you uploaded, not the file on your disk. Without it, next week’s chat reasons over stale notes. - Run a weekly synthesis pass. Once a week ask: “Based on research-log.md, which themes recur across sources? Where are the contradictions, and which perspectives am I missing?” Use the answer to decide what to read next.
The payoff shows up around week three. By then the log holds a dozen extracts, and ChatGPT can flag that the paper you read in week one quietly contradicts the one from week three — a connection you would never have made from memory.
Per-plan limits you will actually hit (June 2026)
The constraint that matters is not your context window; it is how many files a Project will hold and how often you can re-upload. Current figures, per OpenAI:
| Limit | Free | Plus ($20) | Pro / Business / Enterprise |
|---|---|---|---|
| Files per Project | 5 | 25 | 40 |
| Max file size | 512 MB | 512 MB | 512 MB |
| Tokens per text/doc file | 2M | 2M | 2M |
| Files attachable per message | 20 | 20 | 20 |
| Upload rate | ~80 files / 3 hrs | ~80 files / 3 hrs | ~80 files / 3 hrs |
Storage is shared across all chats, Projects, and custom GPTs: 25 GB per user, 100 GB per organization. A single research-log.md will never approach the size cap — the real ceiling is the files-per-Project count, which is why one master log beats one file per source on Free and Plus. Limits change; confirm yours on OpenAI’s File uploads FAQ before a big batch.
Mistakes that quietly break compounding
- Treating each chat as standalone. If you never re-upload the log, every chat starts from zero and the whole system collapses into ordinary note-taking.
- Forgetting to re-upload after a big edit. The Project keeps reasoning over the cached version. If ChatGPT “misses” a source you swear you added, this is almost always why.
- Paraphrasing instead of quoting. A paraphrase loses precision and you cannot re-find the original passage. Capture quotes verbatim with page numbers from the start.
- Letting one log grow to 50 pages with no structure. Past ~20 pages, split by sub-theme (
theme-a-log.md,theme-b-log.md) and keep an index at the top of each. On Free’s 5-file cap, prune ruthlessly instead.
Advanced tips
- Delimit quotes clearly. Use Markdown blockquotes (
>) so ChatGPT never confuses a source’s words with your own commentary during synthesis. - Toggle Project-only memory. Projects added per-project memory controls in 2026 — keep the research Project’s memory isolated so unrelated chats do not bleed into your synthesis.
- Pair with a real PKM tool. ChatGPT is for synthesis and contradiction-spotting; Obsidian or Notion is for permanent storage. Treat the ChatGPT log as a working buffer, not the archive of record.
- Pick the right model per task. Use GPT-5.5 Thinking for the weekly synthesis pass (it holds far more context, ~256K tokens), and Instant for quick single-source extracts.
Copy-ready extraction prompt
Source: [paste content or URL here]
Extract, as Markdown:
1. One-paragraph plain-language summary.
2. Three key claims as direct quotes with page numbers.
3. One or two open questions this source raises.
4. One paragraph: how does this connect to or contradict the rest of research-log.md?
Begin the block with a header line: date, source citation, and a short tag.FAQ
Why not just rely on ChatGPT memory instead of a file? Memory is designed for preferences and recurring context, not structured research notes. It summarizes inconsistently and you cannot audit what it kept. A re-uploaded file is a stable, inspectable source of truth — you always know exactly what ChatGPT is reasoning over.
How big can research-log.md get before ChatGPT loses track? A text file is capped at 2M tokens, so size is rarely the issue. Practical accuracy degrades sooner: past roughly 20 pages of mixed sources, retrieval gets fuzzy. Split by sub-theme before that point, and run synthesis on GPT-5.5 Thinking for its larger context.
Do I need Plus, or is Free enough? Free works for small reviews — its limit is 5 files per Project, so keep everything in one log. Plus raises that to 25 files and gives you GPT-5.5 Thinking for synthesis, which matters once a review runs past a handful of sources.
Why re-upload the file instead of editing it inside ChatGPT? The Project stores the copy you uploaded; edits you make on your own machine do not sync automatically. Re-uploading is the explicit “commit” step that keeps the model’s view current.