Skip to content

rox1694125-bit/magic-memory

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Magic Memory

An AI concept-learning coach that makes you understand and remember.

Most AI explanations optimize one moment: you ask, it explains, you nod — and three days later it's gone. Magic Memory fixes both halves of learning:

  • Understand — it builds intuition before definitions, makes you predict and retrieve instead of passively reading, and catches the exact gap in your mental model.
  • Remember — it turns what you understood into flashcards on a spaced-repetition schedule, so concepts survive the forgetting curve instead of evaporating after the chat.

It's built for learning unfamiliar concepts — abstract mechanisms, programming/framework ideas, architecture tradeoffs, and confusing pairs of terms ("what's the difference between X and Y?").


How it works

A single learning loop:

calibrate → build intuition → predict → reveal → map → mark boundaries
          → make you retrieve → compress into flashcards → spaced review
  • Calibrate first — finds what you already know, why you're learning it, and how deep to go. Crucially, it matches examples to your background (no code snippets thrown at non-programmers).
  • Intuition before formalism — analogy, fable, tiny example, or real case first.
  • Active recall — it stops and makes you answer before revealing. Nodding along isn't learning.
  • Spaced repetition — strong moments become flashcards scheduled by an SM-2 engine.

What's in this repo

File What it is
SKILL.md The coach's instructions — the portable "brain". This is the prompt.
references/methods.md The five teaching modes, gap-label vocabulary, mastery rubric, learning-science rationale.
references/review-protocol.md Flashcard format, grading scale, the spaced-review workflow.
scripts/flashcards.py A standalone SM-2 spaced-repetition engine. Pure Python 3, zero dependencies.

Install & use — works across tools

Magic Memory is deliberately portable: it's just a text instruction file plus a standalone Python script. Anywhere you can give an AI instructions and run Python, it works. Pick your setup:

Claude Code / Claude skills

Drop the folder into your skills directory:

git clone https://github.com/<you>/magic-memory.git ~/.claude/skills/magic-memory

Then just ask to learn something, or invoke /magic-memory. The agent loads SKILL.md automatically and runs the flashcard script for you.

Claude.ai Projects

Paste the contents of SKILL.md into the Project's custom instructions. Add references/*.md if you want the deeper methods available. Run the flashcard script locally when it gives you the commands.

Codex / Cursor / Windsurf / other coding agents

Point the agent at SKILL.md, or copy its contents into your AGENTS.md / system prompt / custom-instructions field. The agent calls scripts/flashcards.py through its own shell — nothing Claude-specific is required.

ChatGPT / Gemini / any chatbot

Paste SKILL.md as a system prompt or custom instruction. The model will coach you through the loop and hand you the flashcards.py commands to run on your own machine for review.

The flashcard engine on its own

It's fully standalone — use it with or without any AI:

python3 scripts/flashcards.py --help
python3 scripts/flashcards.py add --deck rust --front "Q" --back "A" --hook "memory hook"
python3 scripts/flashcards.py due        # what's due today
python3 scripts/flashcards.py grade <id> 5   # self-grade 0–5, auto-reschedules

Cards are stored in ~/.magic-memory/cards.json (override with the MAGIC_MEMORY_FILE env var). Plain JSON — easy to back up or export to Anki later.


💬 Tip: answer by voice

The single cheapest way to get more out of this skill: answer out loud, in your own words, using voice-to-text.

When you speak your answer instead of typing a short fragment, you're doing real active recall and Feynman-style self-explanation — the exact things that build memory. It's also far faster, so you actually explain in full sentences instead of one-word replies (which teach you nothing).

  • If your device has built-in dictation, turn it on.
  • If not, install a speech-input keyboard. The author uses 豆包输入法 (Doubao), whose voice-to-text is excellent — but any good dictation tool works.

This skill is best used as a conversation you talk through, not a wall of text you read.


Requirements

  • Python 3.7+ (standard library only — no pip install needed).

License

MIT — see LICENSE. Free to download, use, modify, and share.

About

An AI concept-learning coach: intuition-first teaching + spaced-repetition flashcards. Understand AND remember. Portable across Claude Code, Codex, and any agent.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages