Cogniloop
Why Cogniloop

ChatGPT and Claude can answer anything.
That's the bug, not the feature.

For actual learning, the tool that gives you answers is the tool that prevents learning. Cogniloop is built on the opposite principle: lock the AI into questioning, lock the UI into explanation, surface progress visibly so the learner can't lie to themselves.

Side-by-side

Cogniloop
ChatGPT
Claude
Default behavior
Asks you to explain. Refuses to answer.
Answers and summarizes by default.
Answers and summarizes by default.
Bypass-resistance
System prompt + UI lock. Can't break out.
Reroll a prompt — you get the answer.
Reroll a prompt — you get the answer.
Concept tracker
Live: weak → shaky → solid → mastered.
None. Buried in chat scrollback.
None. Buried in chat scrollback.
Adaptive difficulty
Explicit 1-5 scale per round.
Improvised tone, not structured.
Improvised tone, not structured.
Coaching report output
Structured: study plan, gaps, journal prompt, markdown export.
Loose prose if you ask for it.
Loose prose if you ask for it.
Streak / accountability
Daily streak, 90-day heatmap, lifetime stats.
Not shown.
Not shown.
Curriculum templates
JEE, NEET, GATE-CSE, MCAT, AP-CS, ML, System Design, Economics — 8 packs pre-loaded.
Cold start every time.
Cold start every time.
Voice in / read-aloud
Web Speech, free, in-app.
Voice paywall on free tier.
No native voice on free.
Math rendering
KaTeX in questions and answers.
MathJax (good).
Plain text by default.
Privacy
Sessions live in browser only. No accounts, no logs.
Stored on OpenAI servers.
Stored on Anthropic servers.
API key required
No.
Free tier has limits + signup.
Free tier has limits + signup.
Cost
$0. Forever.
$0 limited / $20+ for full.
$0 limited / $20+ for full.

Use cases

Pre-exam revision
Night before JEE/NEET/MCAT/GATE — paste a chapter, drill yourself for 20 min. The concept tracker tells you exactly which sub-topics to revisit before sleeping.
Lecture follow-up
Right after class — paste your notes, see if you actually understood or just transcribed. Identify weak concepts before they pile up.
Self-study from textbooks
Reading a chapter? Paste the section, then close the book. If you can explain it well enough that Cogniloop scores you 2-3, you're good. If not, re-read.
Interview prep
ML/CS roles love system-design and concept questions. Use Expert mode on topics like 'distributed consensus' or 'transformer attention' — get adversarial drilling.
Teaching / TA work
Generate question banks for office hours by running sessions in Exam mode and exporting markdown. Or assign specific topics to students and review their session reports.
Long-term retention
Streaks + history make daily reviews stick. Even 10 minutes a day on different topics beats 4-hour cram sessions for retention (spacing effect).

FAQ

Why can't I just tell ChatGPT to be a Socratic tutor?

You can. But you can also un-tell it in one prompt. The friction is wrong: ChatGPT defaults to answering, you have to fight it to drill yourself. Cogniloop defaults to drilling, you can't make it answer. The default is the product.

Claude has Projects with persistent memory now. Doesn't that solve it?

Projects store text across chats. They don't show you a structured concept-mastery map, don't enforce a question-answer-evaluate loop, don't track streaks, don't render a final coaching report. Memory ≠ workflow.

Isn't this just a wrapper?

It's a wrapper in the same sense Linear is a wrapper around a database, Superhuman is a wrapper around IMAP, and Notion is a wrapper around a rich-text editor. The win is product focus and locked behavior — same primitives, different game.

What LLM does it use?

Default: Pollinations.ai's free public endpoint (openai / openai-fast / mistral — no API key needed). With a server-side GROQ_API_KEY set, it uses Groq's Llama 3.3 70B for higher quality and speed. Multi-provider with retry + automatic fallback.

What about hallucinations?

Cogniloop doesn't generate explanations to feed you — it generates questions and grades your explanations. The model's hallucination surface is tiny. If your pasted notes are wrong, you'll be questioned on wrong stuff — but that's a content issue, not a tool issue. For high-stakes accuracy, paste your textbook content as notes.

Why are sessions client-side only?

Privacy + simplicity. No login, no leak, no GDPR. You own your data, it never leaves your browser. The trade-off: clearing browser data deletes your history. Mark sessions to markdown if you want long-term storage.

When should I NOT use Cogniloop?

When you want answers, not learning. When you're cramming raw facts (use Anki). When you need step-by-step solutions to homework (use a tutor or solver). When you need coding execution (use ChatGPT Code Interpreter or Claude with tools).

Try one round. See the difference.

No signup. No keys. Two minutes.

Start a session →