Puzzle lab · Updated 12 September 2026
How memraize puzzles are made
A daily puzzle can look effortless on the screen. Behind it are two separate jobs: making a valid puzzle and deciding whether it is worth somebody's time. This page opens that workshop. It explains the choices, measurements and limits behind all three memraize games, including the parts a program can prove and the parts that still need a person to make a judgment.
memraize is built and maintained by one independent maker in the UK. The games are free, need no account and keep play progress in the player's browser. If a clue is wrong or a puzzle breaks, the feedback button on every page sends the puzzle id and the report to the maintenance queue.
Crossword Maze: designing the reveal
Crossword Maze is not a standard crossword with a curtain drawn over it. Its hidden grid changes the shape of the solve. At the start, the player receives one across clue and one down clue. Solving either answer reveals every clue that crosses it. Those answers reveal more crossings, so the puzzle opens as a chain of discoveries rather than as one large clue list.
The generator first places roughly ten to twelve answers on a sparse 7 × 7 grid. A normal crossword filler likes dense intersections because they make the grid compact. Here, too many intersections are a defect: one answer would expose most of the board and the maze would disappear. The placement score therefore rewards a new word that attaches to one existing answer and extends the chain. It also rejects accidental letter runs and any repeated answer inside the same grid.
Once the grid exists, it becomes a graph. Every answer is a node; every crossing is an edge. Starting from the two visible clues, the checker measures the longest shortest route to the remaining answers. We call that the reveal diameter. A grid with a diameter below the configured floor is discarded and filled again. This small measurement catches the most disappointing maze: a technically valid crossword that unfolds in one move.
Clues come from an owned bank built for each language, not from translating an English finished puzzle. English, Turkish, French, German, Italian, Spanish, Portuguese, Finnish, Polish and Danish each receive a separate grid and clue set. Candidate clues pass answer- leak, safety and wording checks before they enter the bank. The daily difficulty combines clue rank, answer rarity and reveal diameter, with easier ceilings early in the week and harder ones later.
A quality check the first version missed was repetition across days. A large clue bank does not help if a deterministic daily filler keeps choosing the same common answers. The current generator records the previous 28 days for each language and removes those answers from the next day's candidate pool. It also restricts the filler to words that have an approved bank clue. That keeps an automatic fallback from quietly turning an unreviewed word into published content.
Glyphs: proving there is one answer
Glyphs began with a one-sentence rule: place every symbol once in each row and column, without repeating a symbol anywhere on the same diagonal. Black cells stay empty and a diagonal continues through them. Some open cells are also meant to remain blank. The finished board is a set of overlapping placements, one for each symbol.
This is closer to a family of queens puzzles than to a Latin square. A row or column can rule out a symbol nearby; a diagonal can rule it out several cells away. That distance is why the black cells matter visually but do not interrupt the logic. It is also why a vague instruction such as “not diagonally adjacent” would describe a different and much easier game. The rule text, conflict warnings and solver all use the full diagonal.
Generation starts from a completed board. The puzzle maker then removes givens one at a time and asks the solver two questions: does the board still have exactly one solution, and what deductions are needed to reach it? If removing a given creates a second valid board, that removal is undone. This carve-and-check loop continues until the target band is reached. Easy, medium and hard days change the board size, symbol count and score floor instead of merely changing a label after the fact.
The useful part of a uniqueness proof is what it prevents. Without it, a player can obey every visible rule and still be told that a different hidden arrangement was “correct.” That is not a hard puzzle; it is a broken contract. The generated JSON records that the uniqueness check passed, and the web loader checks the dimensions, givens, symbols and blocked cells again before a board can render.
Glyphs uses no vocabulary, so one daily board works in every supported language. The interface text is translated, but the actual reasoning problem is shared. Result sharing encodes only the player's path and outcome, never the hidden solution, so posting a solve cannot spoil the board for somebody who has not played yet.
Nonogram: fitting a picture into 100 cells
A nonogram works in two directions at once. Each number describes a run of filled cells in a row or column; the overlaps between those descriptions settle the picture. At memraize the target is a 10 × 10 board because it stays legible on a narrow phone without turning each cell into a tiny tap target. That limit gives every picture only one hundred bits of ink.
Candidate pictures begin as open-licence icon shapes. They are rendered at the final grid size, not judged from their smooth source artwork. Thin strokes can vanish; two separate parts can merge; a familiar object can become a black lump. Candidates with weak silhouettes, unsuitable subjects or nearly identical shapes are removed from the pool. The picture name stays hidden until the solve, so it cannot act as an extra clue.
The generator derives every row and column clue directly from the chosen bitmap. The web parser independently recreates those runs from the solution and rejects a file if even one clue disagrees. It also verifies the board dimensions and any cells revealed at the start. These are simple checks, but they close a nasty failure mode: one wrong number can make an otherwise attractive picture impossible to complete.
Difficulty comes from the work a beginner-style solver has to do. Full and empty lines settle quickly; large runs create forced overlaps; later sweeps combine facts from rows and columns. The score records how much the first sweep decides and how many sweeps are needed. A few starter cells can be revealed to give a hard picture a fair entrance, but they must agree with the solution and must not give away the silhouette.
The pool rotates before repeating a picture. That matters because a remembered silhouette bypasses the clues entirely. The operations monitor measures how many unused pictures remain and raises a warning months before the pool would cycle, leaving time for a proper curation pass instead of a hurried automatic refill.
What automation cannot decide
A solver can prove uniqueness. A schema can prove that a file has every required field. A filter can catch a clue that contains its own answer. None of those checks can prove that a clue feels natural to a native speaker, that a picture is delightful rather than merely recognizable, or that a puzzle has a satisfying rhythm.
For that reason, publishing and monetising are separate decisions. A generated puzzle may be available to play while it gathers feedback, but automated archive pages are not ad surfaces and are not submitted as search landing pages. Before a generated surface is eligible for ads, its editorial review must be recorded. Until that work is complete, the site uses an account-verification tag rather than loading ad code across every page.
Reports are part of the quality system, not a substitute for it. A player can flag a wrong, offensive or awkward clue from the clue panel. An offensive-content report can remove the affected puzzle immediately; other reports enter a repair queue with the game, language, date and clue number attached. Retired answer words are removed from future fills, so one correction changes the pipeline instead of patching only one old file.
The small details players should not have to notice
Every game works at 320 pixels wide and with a keyboard. Buttons have names for screen readers; colour is never the only signal; motion respects reduced-motion settings. A completed result can be shared without exposing the answer. Future-dated puzzle files can sit safely in the release bank, but their URLs return nothing until the date arrives.
Those details do not make a good puzzle by themselves. They remove distractions around one. The aim is modest: open the page, understand the rule, find a foothold, and have the next few minutes reward careful thought. The measurements exist to protect that feeling, and this page will change when the process changes.
Questions about the process or a particular puzzle are welcome through the contact page.