repaso · Learn and don't forget, without an app

Tell your AI agent what you want to learn, or what you just learned. It makes the cards, quizzes you five minutes a day, grades your answers in your own words, and brings each card back right before you would forget it.

Español · How it was built

A howware: your know-how, put to work by AI. Here the know-how is how people learn for good: small questions, answered in your own words, at growing intervals. The agent does what it does well (making cards, asking, judging an imperfect answer, explaining) and an exact tool does what it does badly (keeping count of thousands of reviews and their dates, with FSRS, the scheduler Anki uses). It needs only the memory, and works for any subject, in any language.

› Repásame.
  Tienes 6 tarjetas hoy (unos 3 min). Racha: 12 días.
  1/6 · ¿Por qué no se delega la responsabilidad final?

› Porque aunque otro haga el trabajo, el que responde por el resultado sigue siendo uno.
  ✓ Bien.
  2/6 · ¿Qué tres partes tiene el feedback SBI?

› Situación, comportamiento… y no me acuerdo de la otra.
  Casi: te faltó Impacto (situación, comportamiento, impacto).
  …
  ✓ Listo: 6 repasadas. Racha: 13 días. 🎉 25 tarjetas dominadas.

What it does

You say	It does
«I want to learn to manage people»	Asks what for, how deep and by when; proposes a study plan in units; you approve it once. Makes the cards of the first unit, and of each next one when it comes up
«Today I read that…»	Turns it into one or two cards, right away, without questions
«Quiz me»	A session of five minutes: what is due first (most at risk of being forgotten), then a few new cards. New cards from the plan are taught first, then asked. One question at a time; you answer in your words
(you answer)	The agent grades it: again, hard, good or easy. On a miss it gives the right answer; on a good answer, as you prefer. What you missed comes back at the end of the session
«How am I doing?»	Streak, mastery per topic and unit, how much you actually remember against your target, the next milestone, the cards you miss most
«For the exam on December 15»	Says if you are on pace, and how many new cards a day the date needs

Habit without guilt. A streak with two freezes a month (a cold shouldn't erase forty days), a small daily goal, mastery you can see per unit, milestones, and after a break a normal five-minute session instead of two hundred cards. Everything shown is computed from your reviews; nothing is invented to keep you hooked.

How it works

Part	Here
Protocol	skill/repaso/SKILL.md: making good cards, the grading rubric, the session, plans, coming back after a break, in plain words
Tool	repaso.py: FSRS (checked against its reference implementation), the session within its minutes, one unit of the plan at a time, the streak and its freezes, milestones, mastery, measured retention
Adapter	The memory keeps topics, plans, cards and every review. Each card's schedule is recomputed from its reviews, so it never drifts and can be audited

Install

git clone https://github.com/howware/long-term-memory-for-AI && python long-term-memory-for-AI/tools/install_skill.py
git clone https://github.com/howware/repaso && python repaso/tools/install_skill.py

Requires Python 3.9+ and project-memory 0.4.0+. For Codex: python repaso/tools/install_skill.py --target ~/.codex/skills/repaso. Then open your agent in a folder of your choice and say: "I want to learn …".

Privacy and limits

Designed to stay on your computer	What the AI agent sees
The memory file: topics, plans, cards, every review, your settings	The cards, your answers and your progress, when it queries them
Material you keep in the folder and don't share	What you type or share in the conversation

What it can't do or guarantee:

• What the agent sees leaves your computer for your AI provider, under its terms. The memory is not encrypted.
• Cards made by the agent can be wrong, above all those from its own knowledge. Each card records where it came from; say "that card is wrong" and it is corrected (and started over if needed).
• A grade is the agent's judgment of your answer, not a measurement. You can dispute it.
• No notifications: the reminder appears when you open a conversation in that folder.
• Text only in this version: no images or audio, no import from Anki.
• Tested in a real conversation on Claude Code only; Codex will be added to howware.json once tested.

Everything is open source: the protocol, the tool and the memory can be read and checked.

Development

git clone https://github.com/howware/long-term-memory-for-AI ../long-term-memory-for-AI
python -m unittest discover -s tests -v
git ls-files -z | xargs -0 python tools/check_private_data.py

The tests run against a temporary memory with synthetic cards, and check FSRS against reference values from py-fsrs (MIT). See AGENTS.md, and how it was built with howware-builder: docs/building.md.

License

Apache 2.0. The FSRS formulas and default parameters come from open-spaced-repetition (MIT).