// the find
nagisanzenin/engram
Evidence-based learning engine for Claude Code — first-principles curricula, free-recall verification with receipts, FSRS-scheduled memory, and explorable artifacts. Learn anything; keep it.
Engram wraps Claude Code (and five other AI coding assistants) in a structured learning loop: Socratic tutoring, blind-graded free recall, and FSRS-4.5 spaced repetition, all in local files. It's for developers who notice that asking Claude to explain something doesn't actually make them remember it a month later. The core insight is that the tutor and the grader are separate agents — one teaches, the other grades your recall without ever seeing the lesson.
The blind assessor architecture is the right call and actually enforced structurally — the grader runs in a separate agent context, sees only the rubric and your words, never the lesson. This prevents the obvious failure mode where the tutor's enthusiasm inflates your self-assessment. The grader audit section is one of the most honest pieces of self-documentation in any open-source project: they found their own gold set was rewarding leniency (the fooled grader scored *higher* than the correct one), documented the circular validation problem explicitly, and left a genuine disagreement in rather than laundering it. The engine itself (engram.py) is stdlib Python with no network code, and that constraint is enforced by an AST-parsing selftest in CI — not a pinky promise, an actual build gate. FSRS-4.5 with personal parameter fitting at 64+ reviews is the current best practice for spaced repetition, and fitting it to your own forgetting curve rather than using population defaults is the right call.
The README is a research paper defending its methodology, not documentation. Install instructions are buried after thousands of words on gold set adjudication statistics and effect sizes. Someone evaluating whether to spend ten minutes on this tool shouldn't have to find the install command by scrolling. All the architectural separation between tutor and assessor lives at the prompt level — if Claude doesn't follow skill instructions precisely in a long session (which happens), the grader might see context it shouldn't, or the tutor might reveal answers early. The whole receipts-prove-mastery claim degrades silently when instruction-following breaks down, and there's no detection for it. The grader-health certificate expires on model swap by their own design, which means every time Anthropic updates Claude Sonnet the audit is void until someone re-runs 258 items manually. For a tool whose core differentiation is 'blind-graded receipts,' tying that guarantee to a specific model version is a fragile foundation. Multi-platform support is genuinely uneven: Antigravity drops subagents entirely, OpenClaw needs an internal-hooks config flag and fires the nudge only on specific commands, and the 'one state folder, every platform' claim is technically true but the experience varies more than the table in the README suggests.