How Recall turns your history into a hint.
No magic. Your submissions go in, plain code does the maths, and the AI only writes the last sentence.
What happens when you sync
Recall reads your solved list and your submission history from inside your own logged-in LeetCode session. It sends one request per second, so LeetCode never gets hammered. A large account takes a few minutes. You can close the tab and come back — it picks up where it stopped.
It keeps the verdict of each submission: passed, wrong answer, timed out, runtime error. For failed ones it also keeps the failing test case and the error message. Your source code is only stored if you turn that on, and it's off by default.
How Recall picks a related problem
Recall ships a map of 116 solving patterns — things like "sliding window you grow and shrink" or "dynamic programming over intervals". Every pattern was checked by hand against real problems.
When you open a problem, Recall finds which patterns it belongs to, then looks for problems you solved that share one. It prefers the closest match:
| Match | What it means | Strength |
|---|---|---|
| Same sub-pattern | You solved something that works the exact same way. | Strongest |
| Same technique | Different problem, same tool. | Good |
| Same topic | Only the tags overlap. A weaker link, and Recall says so. | Weak |
If nothing scores well enough, Recall shows you nothing rather than inventing a connection.
How the four levels work
The level isn't chosen by the AI. It's decided by code, from what you've actually done — never from how you word your message. Asking harder doesn't get you further.
| Level | What you get | How you unlock it |
|---|---|---|
| 1 | The shape of the problem in plain words. No technique names. | Just open it |
| 2 | Names the pattern, maps it onto one problem you solved — as far as the state. | Say what you're thinking |
| 3 | What's wrong with your failing test case. Not the fix. | Submit once |
| 4 | A short skeleton with the key lines blank. | Ask for it |
One exception: if you submit and fail, Recall jumps straight to diagnosing that failure. You shouldn't have to ask twice.
What gets checked before you see a hint
The AI writes the sentence. Then code checks it, and rewrites or rejects anything that breaks the rules:
- Mentions a problem it wasn't given? Removed.
- Contains code below level 4? Stripped.
- Says "that's your only bug"? Deleted — no single change ever finishes a problem.
- Doesn't end with a question? Sent back.
- Too long for the level? Trimmed.
Every hint you receive has been through that. If the AI fails or is unavailable, you get a plain templated hint instead, clearly marked — not a made-up one.
How Recall knows when to hedge
Not all error diagnoses are equally reliable. Recall measured how often each kind of guess was right, and speaks accordingly.
If it's confident — say, an integer overflow — it tells you plainly. If it's less sure, it asks a question instead of stating a cause. It won't tell you something it can't stand behind.
Error diagnosis is tuned for C++ runtime errors. Python and Java verdicts are recognised, but the diagnosis is vaguer.
Try it
Free, on desktop Chrome 116 or newer.