I began Denki with a small frustration: spaced repetition can be an extraordinarily thoughtful way to learn, yet using it can still feel like managing another inbox. Anki proved the value of the method and remains an important tool; I was not interested in pretending otherwise. I wanted to explore a different interface around the same serious work: quieter, more legible, and less eager to turn every review into a performance metric.
The name is simply Denki. It is a general-purpose learning tool, not a Japanese-language product and has no language or cultural affiliation. What matters to me is the idea behind it: a calm, local-first studio where cards, notes, and the rhythm of returning to material can sit together.
Calm is part of the system
A calmer interface is not only a visual preference. Review already asks the learner to tolerate uncertainty: Do I know this? How well? When should I see it again? If the screen adds clutter, ambiguous controls, or constant celebration, it competes with that judgement. Denki instead organises material into classes and decks, supports standard and cloze cards, renders Markdown, and keeps working space close by. There is an on-card scratchpad for a quick diagram or equation, as well as deck-level notes for longer observations.
Those features are deliberately ordinary. The point is not to invent a new ceremony around studying. It is to reduce the distance between encountering a question, working through it, and deciding honestly how well it was recalled.
Make the algorithm visible, not theatrical
Under the surface, Denki implements the FSRS 4.5 scheduling model. Its default target retention is 90 percent, and the setting is adjustable. The interface asks for confidence on a five-point scale, then maps those answers onto FSRS's four canonical grades: one becomes Again, two becomes Hard, three becomes Good, and four or five becomes Easy, with a further stability bonus for a perfect five. That translation is a product decision, not mathematical decoration; it lets the prompt use language that feels natural while keeping the scheduler internally consistent.
After an answer is revealed, Denki shows the interval each confidence choice would produce. Those previews use the learner's current scheduler settings and a fixed, no-fuzz calculation so they remain stable on screen rather than shifting between renders. I like this balance: the learner can see the consequence of a choice without being asked to admire the machinery.
Local-first is a promise with consequences
Cards, decks, preferences, and review history live in IndexedDB through Dexie. There is no required account, hosted database, or analytics tracker, and the app is designed to work offline. That keeps the default relationship unusually direct: the learner creates the material, and the browser stores it.
But local-first does not mean magically safe. Browser storage can be cleared, a device can fail, and local data does not automatically become cross-device sync. Denki asks the browser for persistent-storage protection, offers full JSON export and restore, and gently reminds a learner to make a backup once a collection is substantial and the last export is stale. The restore path also revives date values rather than treating them as strings, because that quiet type error could otherwise make restored cards disappear from due-date queries. Data safety is not a settings-page afterthought; it is part of the product.
Keep the doors open
Denki accepts Anki packages and CSV files because a learning tool should not demand that someone abandon work already done. Standard and cloze material can move in, and an individual deck can move back out as CSV. AI card generation is optional rather than structural: the learner chooses an OpenAI-compatible provider and supplies their own key. The submitted source text goes to that selected provider only when the feature is invoked. Without it, the rest of the studio remains useful and local.
Development has also been openly agent-assisted. The repository contains explicit guidance for coding agents: inspect before changing, verify with types, tests, and production builds, protect destructive actions, and favour maintainable components. Agents accelerated implementation and review; they did not remove the need to decide what Denki should value or to inspect the result.
What remains unresolved
The tradeoffs are real. Local browser storage is private and fast, but less convenient than seamless multi-device sync. Importers must cope with formats they do not control. A five-point confidence scale gives useful nuance, but ultimately compresses into four FSRS grades. Optional AI can save setup time, yet generated cards still require editorial judgement. And a beautiful study surface cannot rescue weak questions or replace the work of retrieval.
I think those tensions are healthier when they remain visible. Denki is not a promise to make memory effortless. It is an attempt to make the effort feel deliberate: fewer distractions, clearer consequences, and enough ownership that a learner can understand where their knowledge lives and when it will return.