nagisanzenin/engram

What it solves

Engram transforms AI assistants from simple explanation engines into a comprehensive learning system for humans. It solves the problem of the "fluency illusion"—where a user feels they understand a concept after reading an AI's explanation, but forgets it shortly after because there was no active testing or scheduled review.

How it works

Engram integrates as a plugin for various agentic platforms (such as Claude Code, OpenCode, and others). It uses a multi-agent loop to ensure long-term retention:

1. Curriculum Architect: Decomposes a topic into a first-principles concept map based on necessity rather than chapter order.
2. The Tutor: Guides the user through concepts using a "generation-first" approach, requiring the user to predict and explain concepts back before receiving the full answer.
3. The Assessor: A separate, "blind" agent that grades the user's written answers without seeing the lesson dialogue, providing objective evidence of mastery.
4. The Scheduler: A deterministic engine (using FSRS-4.5) that schedules reviews of concepts just before the user is likely to forget them.

Who it’s for

Learners who use AI coding assistants or agentic platforms and want to measurably improve their actual knowledge and long-term retention of complex technical topics.

Highlights

• Cross-Platform State: Learning progress and schedules are stored in local JSON files, allowing users to move between different AI platforms while maintaining the same learning history.
• Blind Grading: Prevents the "enthusiasm bias" of the tutor agent by using an independent assessor to verify mastery.
• Evidence-Based Pedagogy: Implements retrieval practice, spaced repetition, and the pretesting effect to maximize memory durability.
• Local-First: The engine has no network code and stores all data locally on the user's machine.
• Interactive Visuals: Generates interactive HTML explorables for "threshold concepts" to facilitate understanding through manipulation.