An AI tutor that makes you do the thinking—and keeps receipts

What it does

Engram is a multi-platform plugin that wraps around your AI assistant—Claude Code, Codex, and others—to turn it into a personal tutor for you, not a memory bank for the agent. You pick a topic, it builds a first-principles curriculum, forces you to predict answers before revealing them, and then a separate, blind “assessor” agent grades your responses against a rubric it never saw during the lesson. A local FSRS-4.5 scheduler books future reviews as plain JSON on disk, nudging you when concepts are due to fade.

The interesting bit

The architecture deliberately splits the tutor from the examiner: the agent that taught you is kept in the dark about your grade, so your mastery is built on verified evidence rather than the tutor’s enthusiasm or your own optimism. The README even notes a live session where the tutor thought the lesson went great while the blind assessor marked most answers only partial—a gap that is the whole point.

Key highlights

• Runs fully offline with no accounts, no cloud, and no network code; state lives as local JSON in ~/.claude/learning/.
• One state folder syncs across six agentic platforms, so you can learn in Claude Code and review in OpenCode.
• FSRS-4.5 fits personal forgetting curves and schedules 2–4 minute review sessions.
• Blind assessor writes every grade to disk as a “receipt”; no receipt, no mastery claim.
• Threshold concepts get interactive HTML artifacts with sliders and prediction gates instead of more text.

Caveats

• The Google Antigravity port is incomplete: session nudges aren’t wired up yet, and two subagents (architect and smith) are dropped by that platform’s strict installer.
• On OpenClaw, session nudges require enabling internal hooks and only fire on /new or /reset, not every session; agents are spawned via sessions_spawn rather than registered directly.

Verdict

If you use AI assistants to learn technical topics but forget everything a week later, this is a rare tool that closes the loop with actual retrieval practice and evidence. If you are looking for persistent agent memory or a chatbot that just explains things while you nod, this is explicitly not that.