MemoForge: Reducing Flashcard Friction

Motivation

Spaced repetition tools such as Anki help with retention, but creating a first batch of reasonably clean cards from technical or academic material is often the step that causes people (including me) to postpone starting. The repetitive parts are: extracting text from a document, deciding on question / answer boundaries, removing obvious duplicates and formatting the result for import.

MemoForge is a small side project to automate those mechanical steps for simple cases. It does not try to replace existing review workflows and it is intentionally minimal.

Current behaviour

High level sequence:

1. Upload a PDF (or provide raw text / a URL).
2. Text is extracted and segmented into blocks (simple boundary heuristics; no heavy parsing).
3. A generation pass proposes candidate question–answer, cloze and term–definition items.
4. Obvious duplicates and very low information items are filtered.
5. A deck file suitable for Anki import is produced.

The process is single‑shot: no in‑app editing UI at this point. Manual review after import is still recommended, especially for specialised domains.

Implementation outline

• Extraction prefers embedded text; a lightweight OCR path is only used when necessary.
• Segmentation keeps paragraph structure; limited overlap is used to avoid context loss at boundaries.
• Generation uses multiple narrow prompts instead of one broad instruction to reduce drift.
• Duplicate detection is based on normalised text hashing.
• Packaging uses a small script to write an Anki‑compatible archive.

The design aim is pragmatic reliability over novelty.

Why not a generic chat interface

Using a general chat model manually for each chunk works but introduces repeated overhead (copying segments, crafting prompts, reformatting, checking for near duplicates). Automating that pipeline removes that overhead; the language model is only one part of the chain.

Privacy / data handling

Uploads are processed transiently. Metrics retained are limited to coarse technical counters (e.g. processing duration) to monitor failures. Source document text is not stored long term.

Explicitly out of scope (for now)

• Scheduling or review logic (left to Anki)
• Conversational tutoring layers
• Complex manual editing interface

Use

If you already maintain card decks manually, this may provide an initial draft you can prune. If you have postponed starting because of the initial formatting effort it may lower that barrier slightly. Feedback on failure cases (unusual layouts, code‑heavy documents, mixed language material) is helpful.

Project site: https://memoforge.app